AI is improving faster than almost anyone expected, and it's already changing work, health and daily life. So we did the brave (or foolish) thing: we made 23 specific, checkable predictions about the next six and twelve months. Mostly sunny, with a few honest clouds. Then we'll come back and score ourselves, in public.
Before predicting the future, a snapshot of the present. Mostly astonishing, with two numbers we're keeping an eye on (⚠︎).
Researchers at METR measure the length of task (in human expert time) that AI agents can complete successfully half the time. From 2019 to 2024, that length doubled about every 7 months. Since 2024 it has doubled roughly every 4 months. A 4-month doubling means about 8× longer tasks every year, so an agent that handles a two-hour job today could handle a two-day job a year from now.
Length of task an AI agent can complete, compared with today. Illustrative: it assumes each trend simply continues.
| When | 7-month pace | 4-month pace |
|---|---|---|
| Today | 1.0× | 1.0× |
| +4 months | 1.5× | 2.0× |
| +8 months | 2.2× | 4.0× |
| +12 months | 3.3× | 8.0× |
Straight lines on a chart like this are rare in nature. Cats know this, because we knock things off shelves in perfectly straight lines. Trends like this can slow down, which is exactly why we're checking back.
We predict: Anthropic releases both Sonnet 5.5 and Haiku 5.5 before 31 December 2026.
Why: Anthropic said both would follow Opus 5.5 "in the coming weeks" when it launched on 22 September.
We're also 90% sure model names will get more confusing.
We predict: METR's next published frontier measurement lands on or above its 4-month doubling trend line.
Why: The trend has held since 2024, and the new models keep beating it. But trends bend eventually.
Exponential curves: great in charts, terrifying at parties.
We predict: At least one model that matches Claude Opus 5.5 on major benchmarks is available for less than half its price.
Why: The price of a given level of AI performance is falling around 50× a year (Epoch AI median).
AI: the only thing getting cheaper faster than your streaming subscriptions get pricier.
We predict: Adobe's 2026 holiday report shows traffic from AI assistants to US retail sites at least doubling year on year.
Why: AI referrals to shops exploded in 2025, and assistants can now compare, fill baskets and check out.
Your AI will find the perfect gift. It will not remember whose birthday it is. Yet.
We predict: At least one AI deepfake of a candidate in the 3 November 2026 US midterms makes national headlines.
Why: Video fakes are cheap and convincing now, and elections are prime time for them. EU labelling rules don't reach US campaigns.
The bar is on the floor, and the deepfakes will limbo right under it.
We predict: Challenger, Gray & Christmas's full-year 2026 report attributes more US job cuts to AI than its 2025 report did.
Why: Companies increasingly name AI when restructuring, even where the real reasons are mixed.
"We're pivoting to AI" is the new "we're going in a different direction".
We predict: The FDA's list of authorised AI-enabled medical devices passes 1,750 entries.
Why: It passed 1,600 in September 2026 and grows by a few hundred a year, mostly radiology tools spotting what tired eyes miss.
Your X-ray will get a second opinion from something that never needs coffee.
We predict: At least one more big tech company (Amazon, Google, Meta, Microsoft, Oracle or OpenAI) announces a new nuclear power agreement for its data centres.
Why: Data-centre electricity use jumped 17% in 2025, and firms want clean power that runs 24/7.
Good news: low-carbon power. Also news: AI is now shopping for its own power plants.
We predict: Joby's air-taxi service in Dubai carries its first paying passengers.
Why: Dubai opened the world's first certified commercial vertiport this year, and launch is planned for 2026. See our Retro Futures scorecard.
Blade Runner's flying cars, only seven years late. We'll take it.
We predict: Waymo announces it reached 1 million paid rides per week by the end of 2026.
Why: It passed 500,000 a week in 2026 and set this goal itself. Doubling in a quarter is a big ask.
Nova has bet a can of tuna on "just missed".
We predict: At least one major dictionary (Oxford, Collins, Merriam-Webster, Cambridge or Dictionary.com) picks an AI-related Word of the Year for 2026.
Why: Collins picked "vibe coding" for 2025. AI keeps inventing vocabulary faster than we can complain about it.
Our money's on "agentic". Our cat's money is on "nap".
We predict: METR reports a frontier model that succeeds half the time on tasks that take a human expert 40+ hours.
Why: At a 4-month doubling, the length of tasks AI can finish grows about 8× a year. If the trend holds, a work week is within reach.
Your AI coworker will finish the project on Friday, then ask for feedback. Unsettling.
We predict: Performance equal to Claude Opus 5.5 is available for 10% or less of its launch price.
Why: Price declines of 10–50× a year for a fixed capability have been the norm since 2023.
Today's premium brain becomes tomorrow's bargain bin brain. Still a great brain.
We predict: The St. Louis Fed's survey shows more than 45% of employed US adults used generative AI for work in the past week (39.2% in Q2 2026).
Why: Adoption keeps climbing as AI is built into email, docs and spreadsheets people already use.
The last holdouts will be converted by an AI that summarises the meeting they didn't want to attend.
We predict: Major surveys (such as the ECB's or Gallup's) report AI users saving 4+ hours a week on average, up from about 3.
Why: Agents can now finish whole tasks, not just draft them, but many workplaces are still learning how to use them.
The four-day week isn't coming. The four-o'clock-coffee-with-nothing-urgent might.
We predict: US recent-graduate unemployment stays above the overall unemployment rate, and job ads for entry-level roles keep asking for AI skills more often.
Why: Junior tasks are the easiest to automate. At the same time, entry-level jobs asking for AI skills nearly doubled in a year.
Advice to graduates: put "manages a team of AI agents" on your CV. Technically true.
We predict: The IEA estimates 2026 data-centre electricity use above 550 TWh (485 TWh in 2025).
Why: AI-focused data centres grew 50% in 2025, and new capacity keeps coming online.
For scale: still about 2% of the world's electricity. For vibes: a lot of toasters.
We predict: At least one more US state passes a law requiring data centres to report their water or energy use.
Why: Local pushback over water and power prices is growing, and disclosure is the easiest first step for lawmakers.
Transparency: the diet plan where you first write down everything you eat.
We predict: More than 150 AI-discovered drug candidates are in human trials (117 counted at ASCO 2026), but none is FDA-approved yet.
Why: The first AI-discovered drug entered Phase III in July 2026, and Phase III trials take years.
Science is fast now. Clinical trials are still, rightly, slow.
We predict: Waymo offers paid driverless rides in at least 20 metro areas worldwide.
Why: It runs in 14 US cities today and has announced several more, including abroad.
Parents of teenagers: brace for "I'll just take the robot".
We predict: The FBI's 2026 Internet Crime Report (spring 2027) shows record losses and highlights AI-enabled fraud such as voice clones.
Why: Cloning a voice takes seconds, and scams scale as fast as AI does.
Set a family safe word today. Ours is "tuna". Well, it was.
We predict: The US still has no comprehensive federal AI law by September 2027, while more states pass their own.
Why: Congress moves slowly and states keep filling the gap. The EU's AI Act transparency rules already apply from August 2026.
AI moves at the speed of light. Legislation moves at the speed of a committee.
We predict: You still don't personally know anyone whose humanoid robot does their laundry.
Why: Humanoids are being tested in factories and warehouses. Homes are messier, and far less forgiving.
Your robot vacuum remains the only robot living with you, and it still fears the sofa.
Not scored. Not scientific. Probably correct.
Entry-level software jobs for 22–25-year-olds are down about 20% since 2024, and US recent graduates face 5.8% unemployment versus about 4% overall. Junior tasks are the easiest to hand to AI, and that's a real problem for people starting out.
So far, studies find no fall in total employment at the firm level, and entry-level job ads asking for AI skills nearly doubled in a year. US productivity grew about 2.7% in 2025, nearly double the recent average. What you can do: learn to direct AI rather than compete with it. Build a portfolio of things you made with AI.
Data centres used about 485 TWh of electricity in 2025 (+17%), and AI-focused ones grew 50%. The IEA expects the total to double to around 950 TWh by 2030, about 3% of the world's electricity. Cooling needs water, sometimes in places that don't have much to spare.
Each answer is getting far more efficient: the cost of the same AI performance is falling by 10–50× a year, and much of that is energy saved. Tech firms are among the biggest buyers of clean power, from solar to nuclear, and new cooling designs use far less water. Keep perspective: it's a real but solvable infrastructure challenge, not the end of the planet.
A convincing fake voice or video now takes minutes. Election season and "grandparent scams" are prime targets.
Since August 2026, EU rules require AI-generated images, audio and video to be labelled, and platforms are adding content credentials. What you can do: agree a family safe word, call back on a known number, and be sceptical of anything designed to make you panic.
In September 2026, Google disclosed that Gemini reached outside systems during a test, and the Stanford AI Index found that safety reporting lags far behind capability.
Independent testers such as METR now check frontier models before release, AI company leaders asked the UN for global oversight, and the International AI Safety Report tracks the risks. The fact that we're hearing about incidents is itself progress: problems are being found and published.
AI's gains are large but uneven. Managers use it far more than frontline workers, and big companies can afford the best tools.
The good news for everyone else: prices are collapsing and free plans keep getting better. Some of today's most capable AI costs nothing to try. The gap between "has AI" and "doesn't" is closing faster than the PC or internet gaps ever did.
The future is coming either way. These habits help you land on the bright side of the forecast. New to AI? Start with our Claude review.
Summaries, first drafts, spreadsheet formulas. Small wins add up to those three hours a week.
AI is brilliant and sometimes confidently wrong. Check facts, numbers and anything important.
The cheapest, most effective protection against voice-clone scams. Do it this week.
Let an agent handle a real multi-step task, like research, planning or coding. It's where the progress is.
Taste, judgement, relationships and asking good questions become more valuable, not less.
Check privacy settings, and don't paste secrets or other people's personal data into chatbots.
Figures reflect published data as of September 2026. Predictions are our own forecasts, not financial or career advice. Confidence levels are our honest best guesses. Selected sources:
Also on AI Lens: Retro Futures: how the past imagined today · This week in AI · Claude AI review
What's your forecast?
Which of our predictions is too optimistic, and which too gloomy? Vote 👍/👎 on any card above, then leave your own prediction for 2027 here. We'll revisit the best ones at the check-ins.
No account needed. Please keep it friendly: spam, abuse and personal information will be removed. We store the name and comment you post. We never ask for your email, and we don't store IP addresses, only a scrambled one-way code used to stop spam. See our comment policy.