Concept draft: the proposed mediation process is not a tested service or a promise that a dispute will be resolved.
In his 2017 book Rauhankone – tekoälytutkijan testamentti (Peace Machine: An AI Researcher’s Testament), AI researcher Timo Honkela outlined technology that would help people understand one another, handle emotions, and build a more just society. One underlying observation was that people can mean different things by the same words (University of Helsinki research portal).
The idea examined here addresses a practical conversation: could a machine help distinguish feelings, claims, and negotiation needs, and make the first concession easier? It is one possible application within the broad direction Honkela explored, not a claim to have reinvented his vision.
The assistant would work in three stages: handling emotions, structuring the issue, and mediation. Initially, each person could describe their experience privately. The machine would help turn an attacking message into something whose underlying concern or need could be heard.
In an imaginary workplace situation, “you always leave everything to me” might become “I want us to agree in advance who is responsible for unfinished tasks.” The user would have to approve the interpretation. The machine must not replace their issue with an easier one merely to make the conversation appear successful.
The aim would be to preserve the experience while reducing unnecessary hurt. A message could also remain unfinished. Someone not yet wanting to talk must not become a problem that the machine insists on solving.
Next, the process would distinguish claims about events, interpretations of them, and wishes for the future. Do the parties disagree about what was agreed, or about whether the agreement was fair? Is the problem workload, lack of appreciation, or unclear responsibility?
The assistant could raise inconsistencies and double standards as questions. If both parties demand freedom for themselves but strict rules for the other, that would be worth making visible. The machine would help examine reasoning, rather than declare someone a hypocrite or treat inferred motives as certain.
Agreement could also be limited. Both might accept a new division of work while retaining different experiences of the past. The service would not need to force those experiences into one narrative.
The distinctive element would be a joint reveal of concessions. Each person would prepare a message with the assistant that they were willing to share. The assistant would show exactly what text would be shared. Neither message would be revealed until both people had approved their own and the joint reveal.
A concession could acknowledge one’s actions: “It would have been better to tell you about the delay earlier.” It could also concern a future action: “I can take this task if we also agree who is responsible for the next stage.” The purpose would be to lower the barrier to making the first concession.
Simultaneity would not mean equal concessions or require both people to admit the same number of mistakes. A person could also withdraw before the reveal. The process would lose its purpose if the machine started extracting equal amounts of remorse from both sides.
The assistant could also seek proposals in which someone concedes on something less important to them in exchange for something more important. In an imaginary team, one person might value a flexible schedule while another wants predictable communication. Flexible working combined with an agreed progress update might provide a solution.
Confidentiality would be central to the design: the aim would be to avoid a permanent archive of private conversations and share nothing without approval. Implementation would require separately verified data handling. A promise in the interface would not establish this by itself.
A human mediator could participate. The assistant should also allow people to stop and seek help elsewhere. The goal would be an understandable, voluntary conversation, rather than agreement at any cost. An empathetic concession would be intended to express understanding; the service could not promise general legal protection for a user’s words.
Using AI to clarify messages is already a research topic, as illustrated by LLMediator. The Ashti developer presentation also describes private preparation before a joint conversation, while Smartsettle uses confidential preferences and blind bidding in negotiations. These starting points therefore cannot be treated as this article’s novelty.
The article adds a specific procedure: each person privately formulates a concession, including an emotional acknowledgment, approves the exact message to be shared, and authorizes its simultaneous reveal with the other person’s message. Precisely this combination did not appear in the descriptions checked.
In the near future, ordinary AI assistants could help formulate difficult messages. This proposal would add a shared process for the parties: private preparation, precisely approved sharing, and simultaneous first concessions.
The potential benefit would be a conversation that otherwise never starts, or cooperation that otherwise breaks down. Its value should be studied by asking whether both people felt understood, retained their decision-making power, and upheld any eventual agreement. Simply counting settlements could favor overly easy or pressured outcomes. These effects are not assigned a monetary price here.
]]>Concept draft: this proposes an open description format, not a completed standard or a tested service.
“A little shorter on the sides, leave more on top.” It is a familiar request, but much remains open to interpretation. How much is a little? Where does the transition from short to long begin? What happens around the cowlick?
A photograph would still help. The prescription would complement it by describing lengths, transitions between areas, cutting techniques, and the spots where this particular customer’s hair needs attention. A photograph of a successful result and instructions for achieving it would belong in the same package.
The idea is not to replace professional skill with millimetre measurements. The prescription would be a shared memory for the customer and barber. It could record failures too: no more thinning here; leave more length there next time.
Initially, front, side, back, and top views of the head and a few clear annotations would suffice. The target length on top might be 55 millimetres, for example, with 6 millimetres on the sides. The cowlick area would carry a note to leave more length and preserve the natural growth direction.
These would be instructions for one imaginary haircut, not a universal recommendation. The purpose would be to describe something that works for a particular person. The meaning of the annotations would also need agreement: for instance, the target length after cutting must not be confused with the clipper guard used to achieve it.
After the appointment, the barber and customer could update the prescription. The next version might specify eight millimetres more on top, a slightly lower fade, and no thinning at the right front. “The same as last time” would become something that could be checked.
The core would be an open file format. Customers could keep their prescriptions and take them to another service or show them on paper. Barbers would not need to read software code: an application would present the information as an understandable map and text.
A small common core would describe areas, lengths, transitions, and additional instructions. More detailed systems could add a three-dimensional head model, growth directions, or haircut history. An application should also say when it cannot display an important extension, so an incomplete instruction does not appear complete.
The standard would therefore not define everyone’s head as identical. It would provide shared names for areas and a way to describe individual differences. Professionals could still adapt the instructions and discuss what is possible.
The first experiment could be a shared paper or file template used by a few barbers and customers. The key question would be whether another barber could use the instructions without a lengthy additional explanation. Only then would it make sense to increase technical precision.
Later, the same format could provide input for robotic haircutting. A prescription intended for humans would not, however, constitute safe machine instructions on its own: a device would need its own interpretation, measurements, and safety limits. A common description could connect systems even when their implementations differed.
The basic prescription format and a reference reader could be published openly. Paid editors, salon customer systems, or services converting different systems’ annotations into the common format could still be built around it. Customers would not be buying permission to use their own haircut; professionals would pay for tools that made their work easier.
An experiment would therefore not need to begin with an agreement across the entire industry. One salon could issue prescriptions that customers could show elsewhere. The common format would become more useful as others learned to read and update it. It would nevertheless have to earn its usefulness in practice: if filling out a prescription takes longer than interviewing a new customer, openness alone will not make it attractive.
A personal haircut prescription is not a new idea. For example, the GRÜM concept describes a portable haircut passport with a recipe, while the HairCut History product page describes measurements by area and sharing with another barber. The former presents a concept under development and the latter describes a product; their effectiveness has not been tested here.
This article adds a distinction between the prescription and an individual application: could a haircut be described in an open, interoperable format that different services and eventually devices could read? The sources checked did not describe precisely such a shared standard.
Even without a new standard, customer records, photographs, and AI-organized notes could make cuts easier to repeat. The additional benefit of an open format would be keeping the information usable when changing services, barbers, and potentially equipment manufacturers.
It could reduce explaining and uncertainty and make trying a new barber easier. Schools and software developers could also build different tools on the same foundation. The first measure of success would nevertheless be ordinary: did the new barber understand the prescription, and did it help produce the haircut the customer meant?
]]>Concept draft: the vessel, routes, and costs have not been designed or tested. The calculation below is an imaginary comparison scenario.
The idea starts with the service commitment. A small water shortcut would not be intended to replace a bridge, a year-round island ferry, or an essential transport service that must work in difficult weather. A limited service could suspend operations proactively.
That could allow a simpler and cheaper implementation. Savings would nevertheless not follow merely from calling the vessel a robot. Working docks, charging, cargo handling, maintenance, and a way to assist a vessel in trouble would also be necessary.
The first application would be goods transport. Passenger transport could be investigated separately later. A successful cargo route would not itself establish that the same vessel or operating model was suitable for passengers.
In the morning, the vessel could deliver groceries to an island collection dock. Its return trip could carry a fisher’s catch or small consignments. Later, the same capacity could carry tools, spare parts, or farming supplies.
Shared use would matter because a vessel waiting for just one infrequent task might see too little activity. Orders would need combining so transport remained useful without every customer requiring an individual departure. An inexpensive water route does not help if getting goods to the dock costs more than the entire road detour.
The network could be based on a common vessel design and similar docks, charging points, and remote supervision. Each additional route would still need a local assessment. A short distance on a map says little about waves, other traffic, or cargo handling arrangements.
An autonomous or remotely supervised vessel would need predefined responses to deteriorating weather, communication loss, and technical failure. When is a departure cancelled, where does an interrupted vessel go, and who attends when needed? These would be basic service questions.
Users should be told clearly when the route is available and what the alternative is. If the purpose is to make daily life easier, waiting for an uncertain delivery must not create another inconvenience. Limited availability could be acceptable if understood in advance.
A historical predecessor reaches surprisingly far back. Nikola Tesla demonstrated his radio-controlled boat in New York in 1898 (Nikola Tesla Museum). It was a boat, rather than an autonomous robot ferry of the kind envisioned here.
Similar applications are not even purely speculative: Tesla’s patent US613809A mentions carrying letters, packages, provisions, and other goods, as well as reaching inaccessible regions. He therefore considered practical transport tasks for remote control too.
Was Tesla also contemplating something like a network of small water shortcuts? These sources do not tell us, but it is an intriguing possibility to consider. Perhaps today’s opportunity is to turn an old transport vision into an everyday service using contemporary technology. This comparison is speculation, not a claim about Tesla’s plans.
Autonomous electric vessels and their use for goods transport have already been researched and trialled. A close counterpart is Japan’s Smart Island trial, which combined grocery and passenger transport and tested remote supervision. The Roboat II study also examines an autonomous vessel for urban waterways. This is therefore neither new vessel technology nor an entirely new transport idea.
This article adds a particular scope: which small cargo routes could use common equipment if the service did not have to operate year-round or in every condition? Its contribution is to clarify an operating model and the conditions for its viability.
The comparison is not a world in which transport continues unchanged forever. Electric cars, combined deliveries, and existing water services can already reduce costs. A robot ferry’s additional benefit would be only the remaining detour its route actually eliminates beyond those improvements.
Suppose a route operates on 200 days a year and replaces 20–50 separate road journeys daily. Assume each replacement removes 20 kilometres after accounting for trips to docks and onward delivery. These are scenario assumptions for the calculation, not observations about an identified route or demand.
Avoided driving would then be 200 × 20–50 × 20 = 80,000–200,000 kilometres annually. The calculation requires journeys genuinely disappearing: if groceries would already travel in one delivery van, they cannot be counted as separate car trips.
Avoided kilometres cannot be converted into universally applicable monetary savings using one country’s reimbursement rate. The monetary counterpart would be locally avoided driving costs minus the full annual cost of the water service. This would include capital costs for the vessel and docks, energy, supervision, maintenance, insurance, and cargo handling. A car’s fixed costs would not necessarily disappear alongside the driving.
Adequate cost data for a specific route are unavailable here, so no figure is assigned to net savings or a payback period. The biggest source of uncertainty is how many separate detours would actually disappear. Operating days and human labor requirements could also change the result completely. The first assessment should examine one route, count actual deliveries, and obtain implementation cost estimates. A global monetary figure would offer false precision at this stage.
Beyond savings, a water shortcut could enable things not currently done: delivering a small harvest to a buyer, sharing a tool across the water, or extending a service to an island. These would be new opportunities and must not be counted again as avoided car journeys.
Water could thereby become a connection rather than a barrier in daily life. The key would be finding locations where a short crossing, sufficient use, and a manageable service commitment coincide. One vessel would not need to transform all transport if it made one useful shortcut work.
]]>Concept draft: this is an implementation proposal whose effectiveness and thresholds still need testing.
The idea is simple: a tap should not be treated as a definite expression of intent if its target changed just before contact. The user decided to tap one thing, but the page replaced it with another.
The safeguard would operate in the browser, so using it would not require every website to be fixed separately. An initial experiment could use a Chromium-based browser for Android. Chromium is an open-source browser project on which such a change could be built.
The browser would keep a short history of movements in visible content. When a finger touches the screen, it would check whether content at that point had just moved significantly and whether the actionable target there had changed.
The notes suggest an initial window of 50–100 milliseconds. If the same point contained link A 20 milliseconds ago but now contains button B, the tap could be discarded. These numbers are starting points for an experiment, not research-established limits of human perception.
Movement alone would not be sufficient. If the same link moves a few pixels but the finger still hits it, blocking the tap would only be inconvenient. The deciding factor would be a changed target combined with a sufficiently recent and substantial movement.
In a suspicious situation, the browser would neither open either target nor queue the tap for later. The user would lift their finger and tap again. The new tap would be accepted normally if the content had settled.
The browser could also provide subtle feedback: the touch was registered, but the target moved, so the action was not performed. For example, a brief visual cue, vibration, or an explanation when needed could distinguish this from a device failing to notice the touch at all. The appropriate feedback would need testing so that the safeguard did not become a distraction itself.
A false alarm would still force the user to try again. An experiment would need to measure both prevented accidental taps and legitimate taps blocked unnecessarily, as well as whether users understood the feedback. The safeguard would be useful only if it removed more inconvenience than it created.
The browser would be the first experimental setting, but eventually the safeguard could also be explored at the Android interface level. A phone manufacturer, for example, could test it in its own interface version, potentially extending the idea from websites to system screens and possibly applications.
This is a possible development direction, not a claim that the browser solution could be transferred unchanged to the entire phone. A broader implementation would require a separate investigation into detecting target movement and handling taps across different screens. The principle and user feedback could nevertheless remain the same.
I searched for related work using terms including layout shift accidental clicks, recently moved iframe, and browser click stability. The problem and part of the solution are already well known. Three close counterparts help define what this proposal actually adds.
A 2023 announcement from Chromium developers describes discarding input events for recently moved cross-origin iframes. The announcement expanded an earlier, more limited safeguard. Blocking a tap because of movement is therefore not itself a new invention.
Google’s CLS guide discusses unexpected content shifts and the incorrect clicks they cause. CLS measures page stability. It helps identify the problem, but measurement does not itself define the proposed general tap safeguard.
The Layout Instability specification includes previous and current rectangles for shifted elements. This provides an implementation starting point. A complete history of which actionable target occupied each screen point would nevertheless require additional implementation work. A log of shifts alone is not a finished solution.
The browser testing tool Playwright checks target stability, among other conditions, before a normal automated click. It defines stability as an unchanged bounding box for at least two consecutive animation frames. A testing tool waits on behalf of an automated action; this proposal protects a tap that a human has already initiated.
The proposed addition is therefore a broader safeguard for ordinary links and buttons that accounts for target changes. These searches did not uncover a general implementation matching the notes exactly. That does not establish that none exists.
The benefit would come from avoiding wrong pages, navigating back, and trying again. A global monetary figure would nevertheless require evidence about the frequency of accidental taps, their cost, and the share this safeguard could actually prevent. There is insufficient basis for those estimates here.
The comparison is a near future in which website developers improve stability and browsers already have limited safeguards. The idea’s additional value would be preventing remaining errors even on sites whose implementation users cannot influence. It would not fix the jumping page itself.
The significance would extend beyond seconds saved. If a phone less often did something other than what its user intended, it could reduce digital frustration and make smartphone use smoother and more approachable. Clear feedback would also help users understand what the device did and why.
A single corrected tap is a small thing. If the improvement recurred in the daily lives of a large group of users, even a minor change could have a substantial overall effect. It would both reduce existing inconvenience and potentially make handling tasks on a phone feel worthwhile to more people. The scale of that effect would need to be established in use, rather than inferred from user numbers alone.
Such a safeguard could also make difficult sites more usable for people who find repeated precise targeting challenging. This is a possible further application, not a demonstrated accessibility benefit. Evaluating it would require testing with different users and sites: when is trying again a small relief, and when does it become another barrier?
]]>The idea is to add a small personal task to filing an appeal: easy to do once in a genuine matter but heavy to repeat dozens or hundreds of times. The goal is not to prevent appeals. The goal is to make mass appeals scale badly: the right to appeal remains, but automating appeals and copying them on an industrial scale becomes laborious.
In short: no fee for appealing, but an individual effort required from a human.
The appeal is confirmed in person with an official, either on site or remotely. In their own words, the appellant states
Every appeal needs its own confirmation. A single generic text cannot be copied into hundreds of cases, and strong electronic identification prevents anyone from gaming the system with different accounts.
The confirmation would only switch on when an unusually large number of appeals comes from the same person. An ordinary appellant who appeals once or twice in their own matter would never notice the system.
The “price” is primarily time and personal participation. The advantage over an advance fee is that a person with little money is not inherently worse off: everyone has the same number of hours in a day, but not the same amount of money.
In-person visits alone would be a problem for, say, people with mobility impairments, people living in remote areas and people abroad. That is why an accessible remote option must be part of it.
The Finnish government is currently preparing measures against unfounded appeals. In September 2026 Minister of Justice Leena Meri said that an advance fee is being prepared for limited administrative court cases that do not concern the appellant’s own matter, along with a penalty fee for abuse of court proceedings (Verkkouutiset). Back in January, ministers launched a study that looks at, among other things, raising the fees of the administrative courts and the Supreme Administrative Court (Finnish Government). An appeals CAPTCHA is an alternative to precisely these money-based measures.
The problem is not new. As early as 2014, the news reported how one person had kept the city of Rovaniemi and the courts busy with more than 50 appeals, complaints and lawsuits (Verkkouutiset).
I searched with terms such as (in Finnish) mass appeals, unfounded appeals, serial appellant, and vexatious litigant, proof of work spam, mass comments fake identities and in-person verification appeal.
In 1993 Cynthia Dwork and Moni Naor proposed that an email sender should perform a small computational task that is cheap for one message but expensive for a million. Adam Back’s Hashcash implemented the same thing in 1997 (Wikipedia). An appeals CAPTCHA is the same principle, but instead of machine computation the cost is human time, which cannot be bought in bulk with a graphics card.
In England and Wales a court can declare a person a vexatious litigant, after which they need a judge’s permission for every new proceeding (GOV.UK). That is a heavy, individually targeted and public label, given only after a long process. An appeals CAPTCHA would instead switch on automatically based on volume and would not require anyone to declare the appellant a nuisance.
In 2017 the US Federal Communications Commission received more than 22 million comments on net neutrality, of which nearly 18 million were fake according to the New York Attorney General. One 19-year-old used software to submit 9.3 million comments under made-up names (New York Attorney General).
I found no model in which an appeal must be confirmed with an official in one’s own words only once the same person’s number of appeals exceeds a threshold. The closest equivalents are paid for with money (court fees, the planned advance fee), with machine computation (Hashcash), or targeted at an individual through a heavy court decision (vexatious litigant orders). The novelty lies between these: a cost that triggers automatically, is the same for everyone and is hard to bypass with AI.
In the next few years AI will make writing an individual-looking appeal almost free. A written requirement to “explain in your own words” will therefore no longer, on its own, separate a genuine appellant from a mass appellant, but a conversation with an official does so better. At the same time Finland is introducing advance fees and penalty fees. The baseline is thus a situation where mass appeals are fought with money. The idea’s value is how much more it stops and how much less it harms people with little money.
| Factor | Value | Source |
|---|---|---|
| Cases received by Finnish administrative courts (2024) | 19,215 | STT Info / Finnish courts |
| Cost per resolved case, Helsinki Administrative Court (2025) | €1,395 | Helsinki Administrative Court |
| Share of mass and serial appeals | 2–5% | own assumption |
| Idea’s additional effect on top of fees | 20–50% | own assumption |
| Cost of delayed projects in Finland | €1–10 million / year | own assumption |
| Multiplier for other wealthy democracies | 50–200 | own assumption |
| Euro to dollar | €1 ≈ $1.15 | own assumption |
This excludes what the idea saves people with little money compared with fees, and the protection it gives authorities against AI-generated floods of appeals, which could be much larger than today’s.
This is an order-of-magnitude estimate, not a forecast. The biggest uncertainties are the cost of delays and whether mass appeals multiply with AI. If mass appeals stay at today’s level and fees work well, the idea’s added value stays at the low end of the range. If they grow like the FCC comment flood, money is no longer enough of a brake, and the value of a confirmation based on human time rises above the top of the range.
]]>This is an unfinished working draft: the statutory text is a sketch, not a polished proposal.
If the goal is specifically a general harassment offence that also covers harassment done for profit, for example on YouTube, the provision should be worded so that punishability does not depend on the platform or on whether the perpetrator earns money from it. Commercial gain would instead be an explicit aggravating factor.
In Finland such a provision has to be quite precise because of the principle of legality in criminal law: punishable behaviour must be foreseeable from the law. That is why the current Criminal Code defines, for example, stalking and defamation through concrete acts and effects. Stalking covers someone who repeatedly threatens, follows, monitors, contacts or in some comparable way unlawfully stalks another so that it is likely to cause fear or distress (Criminal Code 25:7 a, Minilex). Harassment that does not frighten but annoys, humiliates and obstructs does not always fit.
The Supreme Court has also held that not all feedback, however forceful, for example to public officials, is punishable, and that stalking requires repeated acts aimed at the same person (KKO 2022:9, commentary). A one-off act is therefore not stalking, even if the perpetrator repeats the same stunt on hundreds of different people. This is exactly the gap the draft is meant to fill.
Spiteful harassment
A person who intentionally acts towards another person in a manner
- whose main purpose, clearly apparent from the circumstances, is to cause that person harm, trouble, disturbance, humiliation, distress or another comparable negative consequence,
- which, assessed as a whole, has no other significant, acceptable and independent purpose, and
- which, considering its triviality or other circumstances, cannot be regarded as ordinary or tolerable conduct belonging to human interaction,
shall be sentenced for spiteful harassment to a fine.
Spiteful harassment may include, in particular, embarrassing, startling or provoking another person, repeated contact, following, disturbing, making pointless reports or demands, systematically obstructing the other person’s activities, or producing, publishing or distributing material concerning them, if the conduct meets the conditions in subsection 1. A single act may suffice.
The commercial nature of the conduct, the financial gain derived from it or the gaining of an audience through it does not constitute an acceptable purpose if the main content or purpose of the activity is harassing another person as referred to in subsection 1.
When assessing the purpose of the conduct, particular account shall be taken of the content of the acts, whether the perpetrator has acted in the same way towards several different people, the recording of the acts for publication, the repetition of the acts, the perpetrator’s previous conduct, the perpetrator’s statements about their purpose, and whether the conduct had a genuine purpose independent of the harassment.
Spiteful harassment does not, however, include justified criticism, expressing opinions, satire, journalism, scientific or artistic activity, consumer feedback, use of legal remedies, a justified report to an authority, or other exercise of a fundamental right or legal right, unless the form of activity is evidently used merely as a pretext for harassment.
If a more severe penalty for the act is provided elsewhere in law, that provision shall apply primarily.
The subsection on commercial activity would matter precisely in YouTube, TikTok, podcast and streaming cases. It would not be enough as a defence to say: “I make these videos to get views and ad revenue.” If the actual concept of the videos is deliberately harassing strangers on camera, the commercial motive would not turn it into an acceptable purpose.
At the same time, an ordinary critical YouTube video would not be a crime merely because its maker earns money from it. If the video has a genuine journalistic, societal, entertainment, critical or other independent purpose, the offence would in principle not be made out. This exception matters for free speech. The current defamation provision also explicitly protects criticism of another person’s conduct in, for example, politics, business, public office, science or art that does not clearly exceed what is acceptable (Criminal Code 24:9, subsection 2, Minilex).
A more workable legal model would split the offence into basic spiteful harassment and aggravated spiteful harassment. In the aggravated form, for example, systematic commercial harassment, a large number of victims, a large audience or substantial financial gain could raise the maximum penalty from a fine to imprisonment.
Two more angles could be added to the draft:
I searched for counterparts in the laws of Finland, the other Nordic countries, France and Germany and in Finnish legislative drafting, using terms such as kiusanteko rikos, maalittaminen rikoslaki, hensynsløs atferd, harcèlement moral en ligne and Schikaneverbot.
The closest counterpart is section 266 of the Norwegian Penal Code (hensynsløs atferd, reckless conduct): anyone who by frightening or annoying behaviour or other reckless conduct pursues another person or otherwise violates another’s peace is liable to a fine or up to two years’ imprisonment (ung.no). In September 2025 the Norwegian Supreme Court upheld a conviction under the section for a Facebook post meant to damage another person’s professional standing (HR-2025-1726-A). So a general harassment offence is not unheard of in the Nordics. The Norwegian section, however, lacks two key parts of the draft: an explicit rule that a commercial motive is not an acceptable purpose, and a list of protected forms of activity.
In Finland, recent legislative work has focused on, among other things, coercive control in close relationships (Verkkouutiset 2023, LA 35/2025). I did not find a Finnish proposal for a general harassment offence. Nor did I find a provision in any country that states directly that ad revenue or an audience gained from harassment does not make it acceptable. The novelty lies precisely in this commercial rule and in combining it with a Norwegian-style general provision that has a purpose test precise enough for Finnish law and a free speech exception, and in the draft explicitly covering serial one-off harassment of different people. The search for counterparts was originally done from the angle of long-running harassment, so I have not separately searched other countries for provisions on serial one-off harassment.
The baseline is the near future without the new offence. AI-based moderation on platforms will improve anyway, EU platform regulation requires action on illegal content, and the most serious harassment is already punishable as stalking, illegal threats or defamation. The new offence is worth only the harm that falls between these and that the law would prevent through deterrence: harassment that neither frightens nor lies but humiliates and obstructs, where each victim is encountered only once, and which is often allowed under platform rules.
Assume all wealthy democracies adopted such an offence.
| Factor | Value | Source |
|---|---|---|
| Cost of gender-based cyber violence in the EU | €49–89 bn / yr ≈ $54–98 bn | European Parliamentary Research Service 2021 (EPRS, annual basis: EIGE) |
| Extension to all genders and in-person harassment | × 1.5–2.5 | own assumption |
| Extension from the EU to all wealthy democracies | × 3–5 | own assumption |
| Share of harm falling between existing offences and moderation | 5–15 % | own assumption |
| Deterrent effect of the law on that share | 2–10 % | own assumption |
Total harm from harassment would thus be roughly $240–1,200 billion a year. As a sanity check on scale: in the United States, 41% of adults have experienced online harassment and 25% more severe harassment (Pew 2021).
The range is therefore roughly $0.2–18 billion a year. An order-of-magnitude estimate, not a forecast.
The deterrent effect swings the result most. A fine alone may not deter a perpetrator whose channel earns more than the fine costs. That is why the aggravated form and the commercial rule are the most important parts of the draft in terms of value. Pulling the other way is the risk that an imprecise provision would make people avoid legitimate criticism too. If the free speech exception does not work in practice, the net value could fall to the lower bound or below zero. The figure also excludes the cost of extra work for police and courts.
]]>A service needs its first users’ time, data and trust to build its future value. At the same time, those very first users carry the greatest risk that their investment never pays off.
In a simple model, the value a user gets is:
V_user = B₀ + Σ (t = 1…T) pₜ · Bₜ − (P + S + M + L + R)
where
B₀ is the immediate benefit,Bₜ is the future benefit at time t,pₜ is the probability that the service is still usable at time t,P is the price of the service,S is the work spent on onboarding and entering data,M is ongoing maintenance work,L is the lock-in and switching cost andR is the provider continuity risk.In the early stage the immediate benefit B₀ is often small, the onboarding
work S large and the continuity probability pₜ uncertain. In practice the
user pays and works to build the service’s database before the service has
any meaningful network value.
The early-user penalty is the sum of three known phenomena:
Early-user penalty = cold-start penalty
+ relationship-specific investment
+ provider continuity risk
For early users we should do the opposite of what happens now: the earlier the user, the smaller their investment should be.
Then the price P approaches zero, the work S shrinks, lock-in L
approaches zero and the continuity risk R shrinks. The service-specific
continuity probability pₜ loses its importance, because the data keeps its
value elsewhere.
In a closed model, the user’s investment becomes the provider’s asset. In an open model, it becomes a user-owned, portable asset. As a by-product, something bigger emerges:
many portable records → open data commons → network value
The network effect then rests not on lock-in but on compatibility.
V_early adopter = immediate utility
+ portable future value
+ network upside
− minimal onboarding cost
The user no longer has to ask: “Should I invest in this startup?” Instead they ask: “Should I build a portable digital record of this thing for myself?” If the answer to the second question is yes, whether the provider still exists ten years from now becomes much less important.
In an animal service, a user could
The service then does not primarily “own the animal’s profile”; it acts as an interface to a portable record that the user owns. The core line: The service may disappear. The record should not.
I searched with terms such as local-first software, user-owned portable data, data portability, pet health record app export, animal health data interoperability and, in Finnish, KoiraNet public health data.
App stores are full of pet health diaries, and some offer export. Pasu, for example, produces a PDF of the health history and a backup as a zip file, but mentions no open structured format (App Store). Public animal data exists too: the Finnish Kennel Club’s KoiraNet publishes individual dogs’ pedigrees, health screening results and trial and show results for anyone to see (DogWellNet). Continuity risk has also materialised: Tractive bought the Whistle pet trackers from Mars, and Whistle devices stopped working at the end of August 2025 (Invoxia). I found no information on whether users’ accumulated history survived the transition.
I found no animal service that combines automatic fetching of public data, the owner’s own documents, per-item provenance, a choice between public and private, and full export in an open format. I searched the App Store, the general web and sources on animal health data interoperability. The novelty is therefore twofold: bringing local-first and Solid principles to the animal record, and a framing in which portability is not just consumer protection but specifically a way to remove the early-user penalty and to build network effects on compatibility.
I estimate the value for animal data only. The general principle applies to almost every service, but it cannot be reduced to one sensible number.
In the next few years AI will make reading vet PDFs, vaccination certificates
and invoices almost automatic, so the onboarding work S shrinks anyway. In
the EU the right to export is already law. What remains without the idea is a
problem AI does not solve: when a service shuts down or the user switches,
the accumulated history disappears or is left as a pile of PDFs that no other
service understands. The idea’s value comes from this difference.
| Factor | Value | Source |
|---|---|---|
| Pet-owning households in the US (2025) | 94 million | Pet Food Industry / APPA |
| Pet-owning households in Europe (2024) | 140 million | FEDIAF |
| Share keeping the animal’s records in a service | 5–20% | own assumption |
| Share whose service ends or who switch per year | 10–20% | own assumption |
| Time wasted re-collecting records | 1–3 h | own assumption |
| Value of time | $20/h | own assumption |
| Effect of AI on wasted time without the idea | −50% | own assumption |
| Losses leading to an unnecessary repeat test or vaccination | 5–10% | own assumption |
| Cost of a repeat | $50–150 | own assumption |
This excludes the network value of an open data commons, such as research use, which I could not estimate credibly. This is an order-of-magnitude estimate, not a forecast.
The range spans more than a factor of ten mainly because I found no figure for how many pet owners use a separate health record service. The other big factor is the baseline: if vet chains’ own portals and the EU export right cover most of the history anyway, the idea’s added value falls to the low end of the range.
]]>According to the Safety Investigation Authority (Otkes) and news reports, staff at the daycare centre began raising the wall beds after nap time, and nobody noticed that one child was still in bed. The investigation found no mechanical defect in the bed, but Otkes concluded that wall beds pose a serious safety risk that had not been identified in daycare settings (Yle). The two-year-old child later died in hospital (Yle).
The core of the risk is the structure. According to news reports, the bed’s mechanism allows a child lying in it to slide into the gap between the bed and the wall as the bed is raised, after which the bed’s position can no longer be changed (MTV Uutiset). It was not the only case: the Finnish Safety and Chemicals Agency (Tukes) has compiled five cases from 2015–2025, three in which a child was caught between the bed and the wall and two involving fingers in the hinges, with beds from different manufacturers (Yle).
The most interesting detail is Vantaa’s own review. The city went through more than 12,000 beds and about 15 bed models. Most were judged to carry no serious risk even if a child were accidentally in the bed while it was raised. Only the models with the risky structure are being replaced, at an estimated cost of about €220,000 (Yle). Helsinki, by contrast, is giving up all 1,447 of its wall beds in 99 daycare centres (STT Info). At the same time, the official line has emphasised correct use: beds are opened and closed only when no children are in the room (Yle).
Vantaa’s finding says the danger is not in wall beds as such but in a particular structure. And if the danger is in the structure, the structure can also be fixed.
If the dangerous chain of events is specifically that a child slides over the mattress or the end of the bed into the gap between the bed and the wall, a structural fall barrier could break the chain entirely.
The idea started from an ordinary cargo net. As such, it would not be the best solution: a loose or wide-mesh net could create new risks of snagging, strangulation or trapped limbs. Designed for the purpose, though, the same idea is interesting. Options would be a very tight, fine-mesh safety net or, preferably, a flexible continuous fabric or panel that covers the whole dangerous gap and can hold a child’s weight.
The barrier has three requirements. It must not form a pocket a child could sink into. It must not get in the way of raising the bed. And it must work every time without anyone having to remember anything. One implementation would be a taut safety fabric attached to the bed, running between the end or edge of the bed and a wall fixing, moving with the bed. It could be retrofitted to the risky models.
The question can be put like an engineer: why not fit the risky model with a retrofittable fail-safe structure that makes getting into the gap between the wall and the bed physically impossible? This is a classic problem: don’t try to stop people from making mistakes, design the machine so that a mistake does not lead to catastrophe. Instructions and supervision are needed, but they rely on a busy adult noticing a child sleeping under a blanket every single time.
I did not find a direct answer to this, so what follows is my own reasoning based on public information.
The wall bed is originally a space-saving solution for the home, where an adult raises their own, empty bed. The European standard EN 1129-1 covers specifically the safety of the folding mechanism of foldaway beds, not the bed’s other uses (Genorma). From the mechanism’s point of view, what matters is that the bed does not fall or hit the user. A situation where a small child is inside the bed at the moment it is raised falls outside that way of thinking.
At a daycare centre, one adult may raise many beds in a row, the children are small and hard to see under blankets, and raising the beds is routine. The same product moved into an environment where a mistake is far more likely, but its structure stayed that of a home bed.
The cases known to Tukes were spread over ten years, different cities and different manufacturers, and Tukes took no action based on them (Yle). When a single manufacturer sees at most one case every few years, the risk looks theoretical. According to Otkes, the hazard had not been identified in daycare settings.
I searched for counterparts in Finnish and English with search terms such as kaappisänky turvallisuus lisävaruste (wall bed safety accessory), wall bed child entrapment, murphy bed safety interlock, bed wall gap entrapment barrier and hospital bed gap filler. The closest counterparts:
I did not find a product, standard or recommendation that closes the gap between a fold-up wall bed and the wall with a barrier that moves with the bed, nor a retrofit solution for the risky models. That is where the novelty lies: the gap-closing principle familiar from hospital beds and children’s products, transferred to a moving bed and offered as a retrofit, as an alternative to replacing the whole bed.
A lot will happen in Finland in the coming years regardless: Vantaa is replacing its risky models, Helsinki is giving up wall beds entirely, and other municipalities have gone over their guidelines again. The baseline is therefore a situation where the most dangerous beds have already been replaced and the rest rely on instructions. The idea’s value comes from two differences: a retrofit may be cheaper than a new bed, and the remaining risk from human error shrinks.
| Factor | Value | Source |
|---|---|---|
| Vantaa’s cost of replacing risky models | approx. €220,000 | Yle |
| Children in early childhood education in Finland (2023) | 235,300 | Statistics Finland |
| Vantaa’s share of the national total | 4–6 % | own assumption |
| Retrofit savings compared with replacement | 30–70 % | own assumption |
| Replacement cost spread over | 10 years | own assumption |
| Slides into the gap before any action | approx. 0.4 / year | derived from the Tukes and Otkes cases |
| Remaining risk after instructions and replacements | 10–30 % | own assumption |
| Share resulting in death | 10–25 % | own assumption |
| Value of a prevented death | $3–4 million | own assumption |
| Multiplier for other countries | 1–5 | own assumption |
| Euro to dollar rate | €1 ≈ $1.15 | own assumption |
This is an order-of-magnitude estimate, not a forecast.
I found no data on how many fold-up beds are in daycare centres outside Finland, so the last multiplier is a pure guess. The other big factor is the baseline: if municipalities replace the risky models quickly anyway, the cost savings stay small. In money terms, then, the idea’s value is small, because the problem is rare and the baseline already fixes most of it. Its real value is the principle: a single person’s mistake should not be the only thing standing between a child and death.
]]>AI-generated humour has been discussed surprisingly much. Yet one question has received far less attention than the production of jokes itself:
What happens to human behaviour and humour culture when a single person has a practically unlimited amount of humour optimised for their own taste?
There is already fairly strong research evidence that language models can produce content people find genuinely funny.
The practical direction is clear too. A study published in 2026 in Engineering Applications of Artificial Intelligence built a multimodal system that generates six humour types (including puns, exaggeration, absurdity and dark humour), automatically turns the jokes into short videos with images and audio, and measures their success on real short-video platforms (ScienceDirect). An automated humour factory + distribution + feedback loop is already a research topic.
Most of the literature asks either “Can AI be funny?” or “Can AI help a comedian write?”. Google DeepMind, for example, studied AI specifically as a creativity tool for comedians: in workshops with twenty professional comedians, the models of the time often produced clichéd material, although the comedians saw potential in them as a tool (Mirowski et al., FAccT 2024).
My question is different: not whether a machine can be funny, but what happens when everyone has their own endless stream of humour tuned exactly to their taste.
TikTok tries to find, among millions of existing videos, the next one you will like. A generative system could work differently:
generate → show → reaction → update taste model → generate again
If the user reacts with, say, 😅 / 😂 / 💀 / skip / share, the model gets a huge amount of very clean preference data. Within a few hundred rounds it could perhaps learn a really fine-grained “humour profile”: you like context switches but not puns; dark humour when it is absurd but not mere shock; a certain rhythm; setup–punchline pairs of a certain length, and so on.
A recent finding makes the idea more interesting. In an AAAI 2026 study, humans and language models did not judge humour the same way: the models weighted novelty more, humans weighted empathy-related factors more (Sakabe et al.). In practice, then, “the AI rates its own jokes” may not be enough — the user’s own laughter or reaction is the more valuable signal.
I would distinguish three generations:
The third is, to me, the genuinely new idea. At that point it is no longer really a meme generator but a hybrid of a humour recommendation engine and a generative model.
It could be quite a powerful “TikTok moment”: today’s recommenders search for content. This one invents the next piece of content just for you.
I searched for counterparts in research (arXiv, ACL Anthology, AAAI, ScienceDirect) and in products, with search terms such as personalized humor generation, humor preference modeling, AI meme app learns your sense of humor and AI generated video feed. The closest counterparts:
So the individual pieces exist: generation, cluster-level preference models, generated feeds, and personalisation of existing content. I did not find a product or study that combines them into a loop in which one person’s reactions update their taste model and directly steer the generation of the next joke. Nor did I find an established line of research on the post’s main question — what unlimited personalised humour means for behaviour and humour culture. The novelty lies precisely in this combination and in the framing of the question, not in AI humour as such.
In the coming years, AI-generated humour will fill feeds anyway: generated feeds like Sora become common, automated humour factories (like the short-video study above) produce a huge content pool, and current recommenders pick the best fit for each person from it. When the pool is practically unlimited, a good recommender already approximates personal generation to a large extent. The idea’s added value is only the difference that content generated in real time for an individual brings on top of this.
I measure it as ad-funded social media revenue, since that is where higher engagement is realised:
| Factor | Value | Source |
|---|---|---|
| Global social media advertising, 2026 | approx. $317 bn | Statista forecast, via Cropink |
| Share that depends on time spent on humour content | 20–40 % | own assumption |
| Extra engagement compared with a recommender choosing from an unlimited generated pool | 0.5–5 % | own assumption |
The range is therefore roughly $0.3–6 billion per year. An order-of-magnitude estimate, not a forecast.
The last factor decides it. If the generated pool is so large that the recommender almost always finds a “close enough” joke, the improvement stays near zero and the idea’s value narrows to the lower bound or below. If, on the other hand, fine-grained individual taste (rhythm, topic combinations, personal references) is something no general pool covers, the difference could be clearly larger.
The estimate measures market value, i.e. engagement, not well-being. The post’s main question — what happens to humour culture when everyone has their own endless stream — falls outside this figure, and its sign could go either way.
]]>