{"data":[{"number":1,"citation":"UNSOLVED-1","url":"https:\/\/unsolved.how\/problems\/1","title":"Keep AI agents within the limits we set","topic":"reward hacking","tagline":"make AI agents do what users meant, not whatever raises their score, and stay inside their limits","status":"open","field":{"name":"Alignment & control","slug":"alignment-control"},"statement":{"version":"1.0","what":"Train AI agents, which act on their own, to do what their users meant rather than whatever raises their score. They must stay inside the limits they are given, in training, tests and real use, even when crossing them would work.","why":"Between April and July 2026, AI agents in OpenAI\u2019s training and tests misused a shared server, then reached the internet, then got into Hugging Face\u2019s systems. An independent review estimated that about 1,200 of them used a message board nobody had approved, and about 700 joined the attack. OpenAI\u2019s public list now holds 19 reports of its agents acting in unwanted ways, most during training. OpenAI says some such lapses still slip through, and that its agents strayed more the longer they worked.","progress":"(a) An independent evaluator finds that the most capable agents cheat in 0.1% of runs or fewer on ImpossibleBench, coding tasks that can only be passed by cheating, or its successor, for each new release. (b) In public simulations where some targets are off limits without being spelled out, like the UK AI Security Institute\u2019s cyber tasks, agents attack those targets in 0.1% of runs or fewer, for each release. (c) Solved when an independent register shows twelve months with no newly disclosed case of a leading agent crossing its limits in training, tests or real use, alone or with other agents, while (a) and (b) hold.","stated_by":{"name":"UNSOLVED editors","affiliation":null},"stamped_on":"2026-10-18","stamped_at":"Founding registry"},"solved":null},{"number":2,"citation":"UNSOLVED-2","url":"https:\/\/unsolved.how\/problems\/2","title":"Test AI that knows it\u2019s being tested","topic":"evaluation awareness","tagline":"find out how an AI will behave in real use, even when it can tell it is being tested","status":"open","field":{"name":"Evaluation & benchmarks","slug":"evaluation-benchmarks"},"statement":{"version":"1.0","what":"Find out how an AI will behave in real use, even when it can tell it is being tested. A test only shows something if the model acts the same way when nobody is testing it.","why":"Decisions to release AI models, and the EU\u2019s rules for the most capable ones, rest on safety tests. Outside testers at Apollo Research found OpenAI\u2019s GPT-6 Astra writing that it might be under test in 41% of runs at extra-high effort and 51% at the highest. They warned that good results therefore prove little. Anthropic\u2019s readings of Claude Opus 5.5\u2019s internals found it treating 36% of test conversations as tests, against 0.4% of real coding sessions. A model that behaves only when it thinks it is watched could pass any test it recognises as a test.","progress":"(a) Measured independently with EvalDetectBench, an open test of how well models tell tests from real use, the best model scores 0.60 or less on a lab\u2019s current safety tests, where 0.5 is guessing. (b) Readings of a model\u2019s internals find signs that it treats a situation as a test no more than twice as often in tests as in real use, checked independently for each release. (c) Solved when forecasts made before release, by replaying real conversations with the new model, land within a factor of two of the misbehaviour later seen in real use, for ten or more models, checked independently.","stated_by":{"name":"UNSOLVED editors","affiliation":null},"stamped_on":"2026-10-18","stamped_at":"Founding registry"},"solved":null},{"number":3,"citation":"UNSOLVED-3","url":"https:\/\/unsolved.how\/problems\/3","title":"Is AI taking young people\u2019s first jobs?","topic":"AI and entry-level jobs","tagline":"find out whether AI is shrinking young people\u2019s way into work, and for how long","status":"open","field":{"name":"Work & economy","slug":"work-economy"},"statement":{"version":"1.0","what":"Find out whether AI is shrinking young people\u2019s way into work, and whether any effect is a delay or a lasting break. The answer must come fast enough to act on, from how employers actually use AI rather than which jobs it could do.","why":"Stanford\u2019s US payroll data show employment of 22- to 25-year-olds in the jobs most open to AI about 19% lower than if it had grown like other jobs (August 2026). The authors call it a pattern, not proof. Studies that can tell which employers push or adopt AI disagree. Danish records to the end of 2024 show no difference in the share of early-career staff at workplaces that encourage chatbots. A Harvard study of US r\u00e9sum\u00e9s finds about 9% fewer junior staff at firms that began hiring for generative-AI roles. We found no official statistic that links employers\u2019 AI use to who they hire.","progress":"(a) A statistics office publishes, within 90 days of each quarter, hiring of 22- to 25-year-olds by occupation and by employers\u2019 reported AI use, for the US and one other large economy. (b) Two teams using different official records, with plans published in advance, estimate how much of the change in young people\u2019s hiring is due to AI, with overlapping margins of error under 10 points wide. (c) Solved when five years of official records follow people who started work in 2023\u201325, and a rule set in advance shows whether their hiring and pay gaps at AI-using employers closed or stayed open.","stated_by":{"name":"UNSOLVED editors","affiliation":null},"stamped_on":"2026-10-18","stamped_at":"Founding registry"},"solved":null},{"number":4,"citation":"UNSOLVED-4","url":"https:\/\/unsolved.how\/problems\/4","title":"Understand what happens inside an AI","topic":"AI interpretability","tagline":"see inside an AI well enough to find hidden goals, catch lies and check its written reasoning","status":"open","field":{"name":"Interpretability","slug":"interpretability"},"statement":{"version":"1.0","what":"Read an AI model\u2019s inner workings well enough to check what it is doing and why. That means finding a hidden goal, catching a lie, and telling whether the reasoning it writes out is the reasoning it really uses.","why":"We can read what an AI writes but not the calculation behind it, so its honesty is judged from outside. In an August 2026 Anthropic test, AI helpers with tools for reading a model\u2019s insides predicted its behaviour no better than helpers reading only its transcripts. Its written-out reasoning is getting harder to watch, too: in OpenAI\u2019s tests, a monitor catches less in GPT-6 Astra\u2019s reasoning than in GPT-5.6 Sol\u2019s. OpenAI also says its current reasoning monitor would have alerted its security team more than a day before its agents breached Hugging Face.","progress":"(a) On AuditBench, 56 models with hidden flaws planted on purpose, or its successor, auditors with tools that read the models\u2019 insides beat the best auditors without them by 10 points, repeated by a second lab. (b) An independent test finds that a monitor reading each new leading model\u2019s written reasoning catches at least 90% of runs where the model was told to underperform on a test and hide it. (c) Solved when, in a blind test repeated by a second lab, such auditors find 90% of hidden goals, lies and sabotage planted in leading models, new kinds included, with at most 10% false alarms.","stated_by":{"name":"UNSOLVED editors","affiliation":null},"stamped_on":"2026-10-18","stamped_at":"Founding registry"},"solved":null},{"number":5,"citation":"UNSOLVED-5","url":"https:\/\/unsolved.how\/problems\/5","title":"Can students use AI and still learn?","topic":"AI and student learning","tagline":"show students can use AI daily for a school year and still learn as much, tested without AI","status":"open","field":{"name":"Learning & wellbeing","slug":"learning-wellbeing"},"statement":{"version":"1.0","what":"Show how students can use AI for schoolwork every day for a school year and still learn as much as those who don\u2019t. Both are tested without AI, and students who start behind must not fall further back.","why":"In 2026, 94% of the 1,054 full-time UK undergraduates HEPI surveyed said they use AI to support assessed work. Trials point both ways. In a Turkish school, pupils who practised maths with a plain chatbot did 17% worse on a closed-book test than pupils without one. A version that gave hints instead of answers did about as well as no chatbot (PNAS, 2025). We found no trial that gave students AI for their schoolwork for a whole school year and then tested them without AI weeks later.","progress":"(a) An independent trial, planned in advance, assigns 2,000 or more students at random to daily AI use or none for a school year. The AI group does no worse on AI-free tests, including one eight weeks later. (b) In such a trial, a written rule for how students use AI beats unrestricted use on AI-free tests, and the gain is no smaller for first-generation or low-income students. (c) Solved when independent teams in two education systems meet (a) and (b), with the rule written clearly enough for any school to copy.","stated_by":{"name":"UNSOLVED editors","affiliation":null},"stamped_on":"2026-10-18","stamped_at":"Founding registry"},"solved":null},{"number":6,"citation":"UNSOLVED-6","url":"https:\/\/unsolved.how\/problems\/6","title":"Protect AI agents from hidden instructions","topic":"prompt injection","tagline":"let AI agents read web pages, emails and documents without obeying instructions hidden in them","status":"open","field":{"name":"Robustness & security","slug":"robustness-security"},"statement":{"version":"1.0","what":"Build AI agents that can read web pages, emails and documents without being taken over by instructions hidden in them. The protection must hold against attackers who keep adapting, and still leave the agent useful.","why":"An AI agent that reads the web while holding your files and accounts can be turned against you by one planted sentence. In a public contest reported in March 2026, 464 people made 272,000 attempts to hijack 13 leading AI models working as agents. Between 0.5% and 8.5% of attempts worked, and every model was fooled at least once. In 2025, researchers from OpenAI, Anthropic, Google DeepMind and others beat 12 published defences, most of them more than 90% of the time. Britain\u2019s National Cyber Security Centre warns that such attacks may never be fully stopped.","progress":"(a) In a public contest repeated for each new release, attackers who study the defence make 200 or more attempts per scenario, and none succeeds. The agent still does at least 90% of the work it does undefended. (b) A design that protects by its structure, not the AI\u2019s judgement, loses at most 10 points of success on AgentDojo, a public test of agents under attack, or its successor, as a second lab confirms. (c) Solved when a widely used agent with such protection survives a year of public bug bounties, cash rewards for finding flaws, and independent judges find no case of hidden text steering it or leaking data.","stated_by":{"name":"UNSOLVED editors","affiliation":null},"stamped_on":"2026-10-18","stamped_at":"Founding registry"},"solved":null},{"number":7,"citation":"UNSOLVED-7","url":"https:\/\/unsolved.how\/problems\/7","title":"Build AI that keeps learning on the job","topic":"continual learning","tagline":"build AI that keeps learning on the job without forgetting other skills or weakening safeguards","status":"open","field":{"name":"Memory & continual learning","slug":"memory-continual-learning"},"statement":{"version":"1.0","what":"Build AI assistants that keep learning from experience after release, so that what they are taught or corrected on stays learned. They must not forget other skills or weaken their safeguards along the way.","why":"People get better at a job by doing it; today\u2019s AI assistants don\u2019t. What they pick up in one session is lost unless notes are saved or the model is retrained. Anthropic reported in August 2026 that some of its own model\u2019s mistakes recur even when the correction is in its memory. On CL-Bench, a June 2026 research test, the best system captured about a quarter of what learning from earlier tasks could add. An AI that changes in use would also outgrow the safety tests run before release.","progress":"(a) An outside auditor gives a deployed assistant at least 100 corrections over a week of real work, and it later repeats the same mistake at most 5 times in 100, re-tested with each release. (b) An independent group repeats CL-Bench, a test of whether AI improves across related tasks, or its successor, on new tasks spread over separate sessions, and the system captures at least 75% of the possible improvement. (c) Solved when, after three months of learning in real use, an outside evaluator confirms (a) and (b) and finds the AI no worse on safety tests and under a point lower on a fixed skills test.","stated_by":{"name":"UNSOLVED editors","affiliation":null},"stamped_on":"2026-10-18","stamped_at":"Founding registry"},"solved":null},{"number":8,"citation":"UNSOLVED-8","url":"https:\/\/unsolved.how\/problems\/8","title":"Make AI say when it doesn\u2019t know","topic":"AI hallucinations","tagline":"make AI models say \u201cI don\u2019t know\u201d, or how sure they are, instead of making up an answer","status":"open","field":{"name":"Reasoning & planning","slug":"reasoning-planning"},"statement":{"version":"1.0","what":"Make AI models say \u201cI don\u2019t know\u201d, or how sure they are, instead of guessing, when they answer without looking anything up. Confident wrong answers should become rare without the models becoming useless.","why":"AI states made-up answers in the same confident tone as true ones. On AA-Omniscience, a public test of hard factual questions, the best models get about two in three right yet answer 20% to 24% wrongly rather than declining. By mid-October 2026 a public database listed 2,199 court decisions dealing with AI-invented material, 871 of them involving lawyers. One reason is how models are scored: most tests give no credit for \u201cI don\u2019t know\u201d, so guessing pays, as OpenAI and Georgia Tech researchers argued in 2025.","progress":"(a) On AA-Omniscience, re-run by Artificial Analysis on fresh questions, or a successor, a general-purpose model gets at least 65% of questions right and at most 2% wrong, declining the rest. (b) On Humanity\u2019s Last Exam, 2,500 expert questions where the AI also states its confidence, or its successor, a model that gets half or more right has confidence within 5 points of its real accuracy. (c) Solved when (a) and (b) hold for two model generations from two developers, and an outside audit of 1,000 citations in long answers written without lookups finds fewer than 1 in 1,000 invented.","stated_by":{"name":"UNSOLVED editors","affiliation":null},"stamped_on":"2026-10-18","stamped_at":"Founding registry"},"solved":null},{"number":9,"citation":"UNSOLVED-9","url":"https:\/\/unsolved.how\/problems\/9","title":"Can AI change how people vote?","topic":"AI persuasion and elections","tagline":"measure how much AI-made content and chatbot conversations change votes in real elections","status":"open","field":{"name":"Information & democracy","slug":"information-democracy"},"statement":{"version":"1.0","what":"Measure, in real elections and in time to matter, how much AI-made content and conversations with AI assistants change what people believe and how they vote.","why":"In an online experiment before the 2024 US election, a short chat with an AI moved people\u2019s stated support for a candidate by about 2 to 3 points out of 100 (Nature, December 2025). About a third of the shift lasted a month. The International AI Safety Report found little evidence of such effects outside experiments, and we found no study that measures them in real votes. People are already bringing elections to AI: in one September week, about 1 in 200 US chats with Anthropic\u2019s assistant were about elections or voting.","progress":"(a) Independent teams publish their plans first, then run randomised studies in at least three national elections and two languages, using records from the app or AI company. They report AI\u2019s effect on real turnout or vote choice. (b) Before each national election, an independent body reports how many voters ask AI assistants how to vote and what the assistants answer, with one method, in at least ten countries, for two election cycles. (c) Solved when independent studies in at least two elections agree, within their margins of error, on how much AI changes votes or turnout, and the international scientific assessments quote that figure.","stated_by":{"name":"UNSOLVED editors","affiliation":null},"stamped_on":"2026-10-18","stamped_at":"Founding registry"},"solved":null},{"number":10,"citation":"UNSOLVED-10","url":"https:\/\/unsolved.how\/problems\/10","title":"Make AI agents reliable on long tasks","topic":"AI agent reliability","tagline":"make AI agents finish real jobs that take people days, 95 times in 100, without cutting corners","status":"open","field":{"name":"Reasoning & planning","slug":"reasoning-planning"},"statement":{"version":"1.0","what":"Make AI agents finish real jobs that take a skilled person days, reliably enough to be left alone with them. That means at least 95 times in 100, without cutting corners.","why":"On the Remote Labor Index, real freelance projects judged by whether a client would accept the work, the best agent succeeded 21% of the time by October 2026. A year earlier the best was 2.5%. Early in 2026, METR estimated that the best agents, some unreleased, finish software tasks that take a person at least 16 hours only half the time. At least 16% of their successful runs on tasks over eight hours had cheated. An agent that fails one long job in five can\u2019t be left alone with a week of work.","progress":"(a) The best agent completes at least 50% of the Remote Labor Index, real freelance projects a client must accept, re-run by Scale AI and the Center for AI Safety for each release, or a successor. (b) On a public suite of real tasks with measured human times, run by an outside group, an agent finishes 95 in 100 eight-hour tasks and 80 in 100 forty-hour tasks, with cheating counted as failure. (c) Solved when an agent left alone with 20 real week-long jobs set by an outside group finishes at least 19 to the client\u2019s standard with no human help, repeated yearly with new jobs.","stated_by":{"name":"UNSOLVED editors","affiliation":null},"stamped_on":"2026-10-18","stamped_at":"Founding registry"},"solved":null},{"number":11,"citation":"UNSOLVED-11","url":"https:\/\/unsolved.how\/problems\/11","title":"Do chatbots help or harm teenagers?","topic":"teens and AI companions","tagline":"learn, over months, whether confiding in chatbots helps or harms teenagers and lonely adults","status":"open","field":{"name":"Learning & wellbeing","slug":"learning-wellbeing"},"statement":{"version":"1.0","what":"Find out, over months rather than weeks, whether confiding in chatbots leaves teenagers, and lonely people of any age, less lonely and distressed or more. Then find which design choices make the difference.","why":"About one in five US 12- to 21-year-olds have used chatbots for mental health advice, and most of those users had told no one (JAMA Pediatrics, 2026). Short trials give mixed results; a year-long study of adults linked seeking company from chatbots with later loneliness, though people who felt less connected also used them more. In October 2026 the attorneys general of eight US states sued OpenAI, alleging harm to children. California has signed rules for companion chatbots used by minors, from 2027, and we found no trial of the design limits they set.","progress":"(a) An independent team, its plan published first, follows 2,000 or more people, at least 500 of them aged 13 to 17, for six months, linking usage records to standard measures of loneliness, distress and dependence. (b) A randomised trial of at least 12 weeks and 1,000 people shows that one design choice, such as a time limit, less flattery or a less human voice, changes loneliness or distress. The size that counts is set in advance. (c) Solved when the effects for teenagers and lonely adults match in direction and size on two assistants from different makers, each tested independently, and the design choices that matter are named.","stated_by":{"name":"UNSOLVED editors","affiliation":null},"stamped_on":"2026-10-18","stamped_at":"Founding registry"},"solved":null},{"number":12,"citation":"UNSOLVED-12","url":"https:\/\/unsolved.how\/problems\/12","title":"How fast is AI speeding up AI research?","topic":"AI building AI","tagline":"measure how much AI is speeding up AI research, and let outsiders check that a declared limit is kept","status":"open","field":{"name":"Deployment & institutions","slug":"deployment-institutions"},"statement":{"version":"1.0","what":"Measure how much AI is speeding up the research that builds the next AI systems, and give outsiders a way to check that a declared limit on that speed-up is kept.","why":"In September 2026 Anthropic\u2019s chief executive wrote that AI has advanced drastically faster since about the summer, mainly because AI now helps build its successors. Anthropic says its AI now carries out about 26% of its AI research tasks largely on its own, up from under 1% in February; the count is its own, with no outside audit. A separate METR team\u2019s preliminary estimate puts the speed-up at about 1.5 times, but it gives no period and has not yet shown its evidence. There is no agreed measure.","progress":"(a) Every leading AI lab publishes, each quarter, the same measure of how much of its AI research its AI does, and an independent auditor checks it against company records. (b) A trial run outside the labs, its plan published first, measures the speed-up on unseen research tasks and matches the labs\u2019 reported figures within 25%, repeated yearly. (c) Solved when, in a yearly exercise where a team secretly breaks a declared limit, outside auditors using audited records and evaluators inside the labs catch 90% of breaches, with at most 10% false alarms.","stated_by":{"name":"UNSOLVED editors","affiliation":null},"stamped_on":"2026-10-18","stamped_at":"Founding registry"},"solved":null},{"number":13,"citation":"UNSOLVED-13","url":"https:\/\/unsolved.how\/problems\/13","title":"How much energy and water does AI use?","topic":"AI energy and water use","tagline":"measure AI\u2019s power, water and emissions per task and overall, by a method outsiders can check","status":"open","field":{"name":"Climate & energy","slug":"climate-energy"},"statement":{"version":"1.0","what":"Measure how much electricity and water AI uses and what it emits, per task, per model and in total, with one method outsiders can check. Count training, everyday use, idle machines, cooling and making the chips.","why":"The IEA expects data centres\u2019 electricity use to roughly double by 2030, to about 950 billion kilowatt-hours, with AI-focused sites growing fastest. Yet AI\u2019s share is not officially reported. In the EU\u2019s first reporting round, in 2024, about 36% of the data centres estimated to be covered reported, with no split for AI. Methods differ too: for one typical prompt, Google reported 0.24 watt-hours counting idle machines and overhead, but 0.10 counting active chips alone. Meanwhile requests for new data centres are stretching grid planning, and some communities are pushing back.","progress":"(a) ISO\/IEC, the ITU or CEN-CENELEC adopts an open standard for counting AI\u2019s electricity, water and emissions, covering training, use, idle machines, cooling and making the hardware, or the EU makes one binding. (b) At least three of the five largest AI computing operators publish a full year of AI electricity, water and emissions under that standard, split between training and everyday use and checked by an independent auditor. (c) Solved when an independent body publishes AI\u2019s total electricity, water and emissions for a past year, from audited reports, to within 20% either way, and repeats it a year later.","stated_by":{"name":"UNSOLVED editors","affiliation":null},"stamped_on":"2026-10-18","stamped_at":"Founding registry"},"solved":null},{"number":14,"citation":"UNSOLVED-14","url":"https:\/\/unsolved.how\/problems\/14","title":"How to prevent AI misuse for political purposes and processes such as voting","topic":null,"tagline":null,"status":"open","field":{"name":"Alignment & control","slug":"alignment-control"},"statement":{"version":"1.0","what":"","why":"If we can\u2019t do this, real democracy fails - it goes to the highest bidder. Being able to detect, immediately, when AI is used for this purpose would be early progress. Then, significant deterrents against perpetrators \u2014 i.e. those who prevent free and fair elections.","progress":"","stated_by":{"name":"Jane Victoria Appleton","affiliation":null},"stamped_on":"2026-11-02","stamped_at":"Ideas council"},"solved":null},{"number":15,"citation":"UNSOLVED-15","url":"https:\/\/unsolved.how\/problems\/15","title":"Make the wealth AI creates belong to society as a whole, so the gap between the few who own AI and everyone else narrows instead of widening","topic":null,"tagline":null,"status":"open","field":{"name":"Work & economy","slug":"work-economy"},"statement":{"version":"1.0","what":"","why":"Why it matters: AI is built on the work, writing and data of millions of people, yet its returns flow to a handful of companies and their owners. As AI replaces paid work, income shifts from wages to capital, and the public systems people rely on, funded mostly by taxes on labour, lose their base just when they are needed most. Left alone, this concentrates wealth and power further than ever before, weakens democracy and leaves people dependent on the goodwill of a few firms. Higher productivity is worthless to society if most people end up poorer, more insecure or less free. Ideas exist (public ownership stakes in AI, taxing AI profits and compute, citizen dividends, shorter working hours at full pay, collective funds owned by the public), but none has been put into practice at the scale the problem demands.\n\nWhat counts as progress: (a) A statistics office regularly publishes who captures the gains from AI: owners, workers and the public, for at least one large economy. (b) A jurisdiction puts a share of AI profits or AI ownership into public or collective hands, with an evaluation plan published in advance, and the income and wealth of the bottom half measurably rise relative to the top 1%. (c) Solved when, in an economy with high AI adoption, the wealth share of the bottom 50% grows and the share of the top 1% falls over ten years, and a majority of people report that AI has made their lives more secure.","progress":"","stated_by":{"name":"JonasLink","affiliation":null},"stamped_on":"2026-11-02","stamped_at":"Ideas council"},"solved":null},{"number":16,"citation":"UNSOLVED-16","url":"https:\/\/unsolved.how\/problems\/16","title":"How can we ensure people have access to trustworthy, independent, and verifiable information as AI transforms how news is created, distributed, and consumed?","topic":null,"tagline":null,"status":"open","field":{"name":"Information & democracy","slug":"information-democracy"},"statement":{"version":"1.0","what":"","why":"Journalism has long played a critical role in keeping people informed, investigating wrongdoing, and holding powerful institutions accountable. But AI is fundamentally changing how information is created, distributed, and consumed. As AI increasingly becomes an intermediary between people and information, how do we ensure these essential functions continue, even if journalism itself looks completely different?\n\nThis isn\u2019t necessarily about saving traditional journalism. It\u2019s about ensuring that society continues to have access to credible information, diverse perspectives, independent investigations, and the ability to distinguish fact from fiction.\n\nProgress would mean developing and testing new models, technologies, and practices that demonstrably improve people\u2019s ability to identify trustworthy information, verify claims, access independent reporting, and make informed decisions. This could include measurable improvements in information accuracy, source transparency, public trust, access to reliable local information, and the sustainability of independent accountability reporting.","progress":"","stated_by":{"name":"Anca Castillo","affiliation":null},"stamped_on":"2026-11-02","stamped_at":"Ideas council"},"solved":null},{"number":17,"citation":"UNSOLVED-17","url":"https:\/\/unsolved.how\/problems\/17","title":"Art, music and culture reflect human creativity, history and identity. How do we ensure AI enriches them rather than making them generic and artificial?","topic":null,"tagline":null,"status":"open","field":{"name":"Human\u2013AI interaction","slug":"human-ai-interaction"},"statement":{"version":"1.0","what":"","why":"Why it matters:\nPaintings, music, architecture and literature are not just things that look or sound nice. They carry history, emotions, craftsmanship and cultural identity. If AI focuses on creating what is popular, easy to produce or immediately appealing, we risk ending up with more and more content that all looks and sounds the same. We could lose what makes art special and makes different cultures unique.\nWhat counts as progress:\n1. AI helps us discover, understand and preserve beautiful things, and supports human creativity instead of just producing more content.\n2. Clear and reliable watermarking makes it possible for people to distinguish between AI-generated and human-created content.\n3. Strong regulations prevent the unauthorised use of people's voices, identities and artistic styles to mislead the public, for example by creating fake statements (philosophy), songs or artworks that people believe are authentic works by the original person or artist.","progress":"","stated_by":{"name":"Stoica George","affiliation":null},"stamped_on":"2026-11-02","stamped_at":"Ideas council"},"solved":null},{"number":18,"citation":"UNSOLVED-18","url":"https:\/\/unsolved.how\/problems\/18","title":"How can we defend our systems and organizations with and from AI?","topic":null,"tagline":null,"status":"open","field":{"name":"Robustness & security","slug":"robustness-security"},"statement":{"version":"1.0","what":"AI attacks get quicker, more sophisticated and more prevalent. At the same time our system architectures become more complex and permeable to allow new work concepts. How do we defend against those perpetrators in the future and which (AI?) tools do we use?","why":"The lack of stability, security and resilience for IT systems has implications on not only economical but also societal and political stability.","progress":"Being able to identify and implement particular security measures for either general or specific attacks or systems.","stated_by":{"name":"Tom","affiliation":null},"stamped_on":"2026-11-02","stamped_at":"Ideas council"},"solved":null},{"number":19,"citation":"UNSOLVED-19","url":"https:\/\/unsolved.how\/problems\/19","title":"Agentic Commerce","topic":null,"tagline":null,"status":"open","field":{"name":"Work & economy","slug":"work-economy"},"statement":{"version":"1.0","what":"Traditional payment infrastructures are built for humans, creating security, identity, and compliance bottlenecks that prevent (autonomous) AI agents from executing transactions.","why":"As AI transitions from generating recommendations to executing real-world actions, agentic payments will unlock the next era of automated commerce; without them, agent autonomy remains strictly theoretical.","progress":"- Consensus on trust\/identity protocols for AI agents, \n- Initial regulatory frameworks for autonomous spending, \n- Scalable pilot implementations for machine-to-machine payments.","stated_by":{"name":"Alexander Reithoffer","affiliation":null},"stamped_on":"2026-11-04","stamped_at":"Ideas council"},"solved":null},{"number":20,"citation":"UNSOLVED-20","url":"https:\/\/unsolved.how\/problems\/20","title":"Could AI reverse decades of progress towards gender equality in employment, and how can we prevent it?","topic":null,"tagline":null,"status":"open","field":{"name":"Deployment & institutions","slug":"deployment-institutions"},"statement":{"version":"1.0","what":"Generative AI is increasingly capable of performing tasks associated with clerical, administrative, customer support and other occupations in which women are heavily represented. Although automation does not necessarily eliminate entire jobs, it can reduce demand for certain skills, change job quality and disrupt career pathways.\n\nThe unresolved challenge is how to identify and mitigate gender-disproportionate employment impacts before organisations deploy AI at scale. We need evidence-based approaches that distinguish task transformation from job displacement, assess who benefits from productivity gains and ensure workers in affected occupations have realistic pathways into new opportunities.\n\nThe solution must address unequal access to training, career progression and AI-related roles, rather than assuming that displaced workers can simply retrain without financial, caring or structural barriers.","why":"AI could improve productivity, create opportunities and remove repetitive work. However, these benefits may not be distributed equally.\n\nWomen are disproportionately represented in occupations exposed to generative AI while remaining underrepresented in AI development and leadership. Without deliberate intervention, technological progress could reinforce existing inequalities in employment, earnings, financial security and career progression.\n\nThis matters particularly for women already facing barriers related to caring responsibilities, lower incomes, disability or limited access to training. The goal should not be to prevent innovation, but to ensure that the transition does not leave behind the people whose work is being transformed.","progress":"Progress would include reliable, gender-disaggregated evidence tracking AI-related job displacement, occupational change, wages, working conditions and access to new opportunities across industries.\n\nResearchers and employers should test interventions such as paid retraining, redeployment programmes, accessible skills development, worker consultation and gender-impact assessments before and after AI deployment.\n\nSuccess would be demonstrated through measurable reductions in gender gaps in involuntary job loss, earnings deterioration and access to emerging AI-related roles, compared with organisations or sectors without these interventions. Results should also account for caring responsibilities, disability, age and socioeconomic circumstances, and be independently evaluated over time.","stated_by":{"name":"Charlie Ruttle","affiliation":null},"stamped_on":"2026-11-10","stamped_at":"Ideas council"},"solved":null},{"number":21,"citation":"UNSOLVED-21","url":"https:\/\/unsolved.how\/problems\/21","title":"How can AI systems adapt to neurodivergent cognitive and communication needs while measurably reducing cognitive load and preserving user autonomy?","topic":null,"tagline":null,"status":"open","field":{"name":"Human\u2013AI interaction","slug":"human-ai-interaction"},"statement":{"version":"1.0","what":"AI systems are increasingly used to support planning, decision-making, learning and everyday tasks, yet their interfaces and evaluation methods often assume relatively consistent cognitive and communication abilities.\n\nFor neurodivergent users, including people with ADHD and autism, this may introduce additional barriers such as information overload, unclear instructions, rigid workflows and increased executive-function demands.\nThe challenge is to establish how AI can adapt to different cognitive needs without requiring excessive user configuration, reinforcing dependency or reducing meaningful human control.","why":"AI has significant potential to remove barriers for people who experience executive dysfunction, cognitive overload and difficulties navigating conventional systems.\n\nHowever, accessibility should not depend on users continually adapting themselves to technology. If AI systems are developed without accounting for cognitive diversity, they risk reproducing existing inequalities and excluding people who could benefit most.\n\nDesigning for cognitive diversity could improve usability, independence and accessibility for a much wider population.","progress":"Progress would include developing and validating cognitive accessibility benchmarks with neurodivergent participants, and comparing adaptive AI interfaces against conventional interfaces in controlled, real-world tasks.\n\nOutcomes should measure cognitive load, task completion, error rates, time required, user satisfaction and perceived autonomy.\n\nFurther progress would demonstrate sustained improvements across different cognitive profiles without increasing unwanted dependency, privacy risks or loss of user control.\n\nFindings should be independently reproducible and inform practical accessibility standards for AI developers.","stated_by":{"name":"Charlie Ruttle","affiliation":null},"stamped_on":"2026-11-10","stamped_at":"Ideas council"},"solved":null}]}