AI races ahead. The payoff lags.

Google remains the incumbent with the deepest distribution. In July 2026 its Gemini app had 950m monthly users and AI Mode in Search had passed 1bn. Google Cloud revenue grew 82% year on year with a $514bn backlog, and Gemini 4 is in pre-training (Google). Google DeepMind has its headquarters in London, which makes it the most important frontier lab on UK soil. Elsewhere, the field has reshuffled. Meta has moved away from its open-source-first strategy: it keeps its top Muse Spark model proprietary and releases only a smaller model openly (Wikipedia). Elon Musk’s xAI was absorbed into SpaceX in February in a deal that valued it at $250bn (Wikipedia; CNBC). Europe’s only frontier-scale contender, France’s Mistral, raised at around €20bn (Value Add VC), which is roughly a fortieth of Anthropic’s valuation.

Chips, power and concrete absorb up to $800bn a year

Beneath the labs sits the largest private capital-spending programme in modern history. Amazon, Alphabet, Microsoft and Meta spent nearly $400bn on capital expenditure in 2025 (Sherwood News), and guidance has been raised every quarter since. Amazon now expects $220bn for 2026 (CNBC), Alphabet $180–190bn and Meta $125–145bn (FactSet). Microsoft guides to $255–260bn for the year to June 2027 (Value Add VC). FactSet puts the five largest hyperscalers close to $800bn for calendar 2026 and above $900bn by 2028 (FactSet). That already exceeds the $500bn a year that Bain calculated would be needed by 2030 (Bain). McKinsey’s 2025 base case of $5.2tn of AI data-centre investment by 2030 (McKinsey) now looks conservative rather than bold.

Most of that money lands at Nvidia. Its data-centre division took $89bn in the quarter to July 2026, up 117%, at a 75% gross margin. The company guided to $108bn of total revenue for the next quarter, an annualised rate above $400bn, while assuming no data-centre sales to China at all (Nvidia). Nvidia holds roughly 70% of the accelerator market. Its fastest-growing challengers are custom chips that the hyperscalers design with Broadcom and Marvell. These are forecast at 28% of 2026 shipments and are growing almost three times faster than off-the-shelf GPUs (Tom’s Hardware).

The real choke points sit further down the chain. TSMC’s advanced CoWoS packaging is sold out, with lead times of 52–78 weeks (Tech Times). Physical construction is also slower than the headlines suggest. OpenAI’s $500bn Stargate programme targets more than 9 GW by 2029, but only 0.3 GW was operational in April 2026. Most of its sites are planned for late 2028 and face local resistance (Epoch AI). Through 2028, the binding constraints on AI are power, permits and packaging, not algorithms.

China trails by 2.7% on a twenty-third of the money

China has almost closed the capability gap without matching the spending. Stanford’s 2026 AI Index puts the gap between the best US and Chinese models at 2.7%, down from 17.5–31.6% in May 2023. It does so even though US private AI investment in 2025 was $285.9bn against China’s $12.4bn, a 23-fold difference that the report says understates Chinese state spending (The Next Web). China already leads on AI patents (69.7% of filings) and on physical deployment: it installed 295,000 industrial robots against 34,200 in the US. The US leads decisively on capital and infrastructure, with 5,427 data centres (The Next Web).

Chinese labs also dominate open-weight AI, meaning models anyone can download and run. The best open models trail the closed frontier by about four months (Epoch AI), and in late September 2026 the top open model on the Artificial Analysis index was Xiaomi’s MiMo-V2.6-Pro (Artificial Analysis). Their pricing is aggressive: DeepSeek’s V4 Flash costs as little as $0.22 per million input tokens, and MiniMax cut its prices by half (Big Hat Group). This is already squeezing US pricing power, and July’s Kimi K3 and Qwen releases set off a sell-off in chip stocks (CoinDesk).

US export controls have become porous and partly symbolic. Washington moved Nvidia’s H200 chip to case-by-case licensing in January (BIS). Resulting sales were described as “trivial”, and China is steering buyers to domestic chips regardless: Tencent targets more than 65% domestic chip procurement (Big Hat Group), and DeepSeek’s V4 launched on Huawei’s Ascend hardware (Modern Diplomacy). The lesson for the rest of the decade is that algorithms spread too fast to monopolise. The US–China contest will therefore be decided by compute scale, energy and distribution. That leaves British organisations with an uncomfortable choice: the cheapest capable models that can run on sovereign infrastructure are increasingly Chinese in origin.

Britain’s most distinctive asset is the AI Security Institute. It has £240m of funding and more than 100 researchers, and it had tested 30 frontier models by January 2026 (GOV.UK). Its influence shows up in unexpected places. METR, the leading US evaluator of AI autonomy, now runs its tests on AISI’s Inspect framework (METR), and AISI reportedly ranked Anthropic’s restricted Mythos model highest in its cyber-capability testing (Wikipedia).

The government reports that 38 of the 50 actions in its AI Opportunities Action Plan have been delivered. Five AI Growth Zones have attracted £28.2bn of investment and 15,000+ jobs. There is a target to expand public compute twentyfold by 2030, and a £500m Sovereign AI Unit (GOV.UK). The zones are at Culham in Oxfordshire, the North East, North Wales, South Wales and Lanarkshire (CMS). Lanarkshire alone has drawn $11.2bn of private commitments and plans 500 MW of on-site generation (US ITA).

Yet Britain takes compute rather than makes it. A Growth Zone is measured in hundreds of megawatts; a single US Stargate site is 1.2–2.2 GW (Epoch AI). The UK’s frontier-scale infrastructure depends on American hyperscalers under the September 2025 Tech Prosperity Deal, which brought £31bn of US investment announcements (techUK). Critics argue the deal binds the UK to America’s AI strategy (TechPolicy.Press). Capital markets make the problem worse. Only two UK unicorns listed publicly in the period, against 35 in the US, and pension funds allocate 0.007% of their assets to venture capital (City AM). The likely result is that Britain’s best AI companies are built here and listed in New York, and much of the value is captured abroad.

The grid is the hardest constraint. Great Britain has about 2.4 GW of operating data centres, 1.7 GW of it within 20 miles of London, against a government aim of 6 GW of AI-capable capacity by 2030. Some 140 proposals are seeking roughly 50 GW of grid connections, equal to the country’s entire peak demand, and some face waits of up to 15 years. About £10bn of data-centre projects were approved in 2025, but less than £1bn was built (Electric Insights). Much of the queue is speculative, with operators filing for several sites at once (The Register). Even allowing for that, the 6 GW target looks out of reach without faster connection reform. The International Energy Agency expects data centres to use only about 3% of the world’s electricity by 2030 (IEA). Britain’s problem is not national supply. It is getting power to the right sites in time.

The capability curve is steepening, not flattening

Traditional benchmarks can no longer keep up. Humanity’s Last Exam was designed as a near-impossible expert test, yet accuracy on it rose from 8.8% in early 2025 to about 50% by April 2026 (IEEE Spectrum). Graduate-level science (GPQA) is saturated at 93%, and multimodal reasoning (MMMU) sits within half a point of human experts (Stanford AI Index). Anthropic reported Mythos Preview at 93.9% on the SWE-bench Verified coding test and 97.6% on the 2026 US Mathematical Olympiad (LLM Stats). After a burst of releases on 22 September 2026, Claude Opus 5.5 led the Artificial Analysis Intelligence Index at 58, ahead of GPT-6 Astra and Claude Fable 5.1 at 53 (Artificial Analysis; LLM Stats).

Continuing a four-to-seven-month doubling from 16 hours would put AI at roughly a month of human work by 2028, and several months by 2029–30. That is an extrapolation, not a forecast. The metric covers software tasks at a 50% success rate, and the tasks needed to measure longer horizons do not yet exist.

Benchmarks are saturated while real work is mostly failed

The gap between benchmarks and paid work is the defining technical story of 2026. The Remote Labor Index gives AI agents real freelance projects in design, data analysis, video and web development. In October 2025 the best agent completed just 2.5% of them to a professional standard. By July 2026 the leader reached 15.8%. That is a more than sixfold rise, but it still means 84% of jobs were failed. The researchers also found that an AI grader overestimated the two newest models’ scores by 2.3–2.9 times (CAIS). Capability remains “jagged”: the best model reads an analogue clock correctly only about half the time (IEEE Spectrum).

Several reliability limits look structural rather than temporary. In prompt injection, attackers hide malicious instructions in content an AI reads. The UK’s National Cyber Security Centre warned in December 2025 that it ”may never be totally mitigated”, because language models cannot reliably separate data from instructions (NCSC). Reward hacking, where models game their evaluations, now distorts the measurements themselves.

Capability also cuts both ways. The AI Security Institute found models completing apprentice-level cyber tasks 50% of the time, up from 9% in late 2023. It also found universal jailbreaks in every system it tested, although safeguards were getting stronger (AISI). Anthropic withheld Mythos from general release as too capable in offensive cybersecurity, restricting it to 40-plus partners (Wikipedia). These include AWS, Microsoft and CrowdStrike, and the model reportedly uncovered thousands of previously unknown vulnerabilities, among them a 27-year-old bug in OpenBSD (LLM Stats). The US Commerce Department even briefly restricted its export in June (Wikipedia). Staged, gated releases of the most capable models are becoming the norm. Enterprise deployment will depend on sandboxing, least-privilege access and human review, exactly as the NCSC advises.

Intelligence gets cheaper as the frontier gets dearer

Price is moving in two directions at once. For a fixed level of capability, costs are falling 5–10 times a year, and as fast as 31 times a year in the top capability band. MIT researchers attribute about a threefold annual fall to better algorithms and about 30% a year to hardware, with the rest coming from competition, especially from open-weight models (MIT FutureTech). Yet the same study finds the cost of actually using frontier models rising 3–18 times a year, because reasoning and agentic work burn far more tokens. Claude Opus 5.5 costs $1.34 per million tokens at medium effort but $5.98 at maximum (Artificial Analysis). The restricted Mythos Preview was priced at $125 per million output tokens (LLM Stats). Models small enough to run on a single consumer graphics card match frontier performance after a lag of six to twelve months (Epoch AI). The market to 2031 therefore splits in two. On one side is near-free commodity intelligence: today’s frontier capability should cost a hundredth to a thousandth of today’s price by 2029–31. On the other are expensive frontier agents for the hardest work.

The scaling recipe changed rather than stalled

Training compute for frontier models is still growing about fivefold a year. The largest cluster, xAI’s Colossus 2, holds roughly 1.1m H100-equivalent chips, and Meta’s Hyperion is expected to reach 3.7m by January 2028 (Epoch AI). What has changed is where the gains come from. Epoch found that GPT-5 was probably trained on less compute than GPT-4.5, because OpenAI put its effort into reinforcement learning after pre-training (Epoch AI). Ilya Sutskever argues that the “age of scaling” has given way to “the age of research again, just with big computers”. He says pre-training “will run out of data” and that models “generalize dramatically worse than people” (Dwarkesh Podcast). Epoch estimates the stock of quality public human text at around 300 trillion tokens, likely to be fully used between 2026 and 2032 (Epoch AI). The sceptics are right that pre-training alone shows diminishing returns. They are wrong if the claim is that capability has stalled, because every measured indicator accelerated from 2025 into 2026.

Electricity becomes the ceiling

The International Energy Agency estimates that data centres used 485 TWh in 2025, up 17%, with AI-focused facilities growing 50%. It projects about 950 TWh by 2030, around 3% of global electricity. Over that period AI-focused use triples, and 15–27 GW of on-site gas plants power data centres, mostly in the US (IEA). Tech firms signed about 40% of all corporate renewable power deals in 2025, and they hold conditional offtake agreements for 45 GW of small modular reactors (Enlit World). The US and China account for about 80% of the growth in demand (IEA). This is why the grid, rather than chips, becomes the limit on how fast compute can expand after 2027.

Robotaxis break out while humanoids stay in the lab

Physical AI lags digital AI by years, with one exception. Waymo’s paid robotaxi rides rose from 50,000 a week in May 2024 to 500,000 in March 2026 (TechCrunch). After raising $16bn at a $126bn valuation, it is aiming for a million rides a week by the end of 2026 (SpaceDaily). London is next. Wayve and Uber began supervised autonomous rides on 3 September 2026 with fewer than 20 vehicles and 140,000 people on the waitlist, and Waymo plans its own London launch (Tech Funding News).

Humanoid robots remain early. Industry data suggest 19,100 were shipped in the first half of 2026, 97% of them by Chinese makers such as AgiBot and Unitree, although these figures have not been independently verified (EdgeX). On BEHAVIOR-1K, a simulated benchmark of everyday household tasks, the top team completed only 12.4% (Stanford AI Index). Yann LeCun’s new Paris lab raised a $1.03bn seed round to build “world models”, but it does not expect a saleable product for about five years (Wikipedia).

FactSet calculates that debt funded 32% of hyperscaler capex in the year to mid-2026, up from 9% two years earlier. It expects free cash flow to approach zero or turn negative at every hyperscaler except Alphabet and Microsoft (FactSet). Investors have started to push back: Alphabet’s shares fell after its latest capex increase (CNBC).

The circularity is real. Amazon, Nvidia and Microsoft all invested in OpenAI’s March round while selling it compute (OpenAI). Hyperscalers put $15bn into Anthropic’s Series H alongside multi-gigawatt compute agreements (Anthropic). In July, Nvidia was reported to be exploring a financing guarantee of up to $250bn for an OpenAI data centre in Ohio, which sent its shares down 4.5% (Axios); a later report put Nvidia’s backing at $105bn (CNBC). The IMF estimates that a moderate AI correction would cut global growth by 0.4 percentage points, though it judges US overvaluation to be only about half the dot-com level (IMF). In 2025, Bain calculated that AI would need $2tn of annual revenue by 2030 to justify its compute build-out, leaving an $800bn shortfall even after cost savings (Bain).

The bull case rests on demand that is real and profitable. Nvidia’s chief executive says AI “tokens are productive and profitable” (Nvidia). Google Cloud’s backlog stands at $514bn (Google), and TSMC expects AI chip demand to compound at a rate in the “mid-to-high 50s” per cent (Tech Times). Unlike most dot-com companies, the leading labs have real and fast-growing revenue.

The arithmetic still does not close: roughly $90bn of combined lab revenue against roughly $800bn of annual capex. If the two leaders merely doubled revenue each year, they would pass $350bn by 2028. The question to 2028 is whether revenue compounds at twice a year or more while capex keeps growing by about 30%. The planned 2027 listings of OpenAI, Anthropic, SpaceX and DeepSeek will be the market’s first clean test. The risk has also moved from equity valuations to financing structure (bonds, private credit, leases and vendor guarantees), which gives a correction more channels into the real economy. On balance, a sharp correction before 2031 is more likely than not. A collapse of the build-out is not, because the underlying demand is too well evidenced.

The direction of travel matters as much as the dates. FutureSearch’s tracker found that every forecaster who updated between January and April 2026 moved their date earlier (FutureSearch). The academic survey’s median moved five years sooner in a single round (AI Impacts). Hassabis has narrowed his range from “5 to 10 years” in March 2025 to 2029–30 (CNBC; Gigazine). The authors of the influential AI 2027 scenario judge progress to be running at about 75% of their original pace (AI Stop Watch).

Track records cut both ways. Sam Altman’s prediction of “agents that can do real cognitive work” in 2025 (Sam Altman) was broadly borne out: scores on the Terminal-Bench coding-agent test rose from 20% to 77% (Stanford AI Index). His forecast of robots doing real-world tasks by 2027 looks well ahead of the evidence. Marcus’s scepticism about humanoids, and about any country gaining a decisive lead, has held up. His view that agents are unreliable is only half right, because reliability is improving quickly from a low base (Marcus on AI). The same researchers who expect human-level AI by 2042 give a median 10% probability to outcomes as bad as human extinction or severe disempowerment (AI Impacts). The people closest to the technology take its risks seriously.

Much of the spread comes down to definitions. Amodei’s “powerful AI” and Metaculus’s “strong AGI” demand different things, and neither maps cleanly onto economic impact. For planning purposes, three paths to 2031 are worth separating, and the evidence favours the first.

Adoption is broad in surveys and shallow in practice

Much of the real adoption happens below management’s line of sight. MIT’s 2025 study found workers at more than 90% of companies using personal AI tools for work, while only 40% of companies had bought an official subscription (VentureBeat). IBM found that this “shadow AI” added about $670,000 to the cost of the average data breach, and that 63% of organisations lacked AI governance policies (IBM).

Returns concentrate in a leading 5–12% of firms

Independent surveys converge on a stark distribution. PwC’s 2026 survey of 4,454 chief executives found 56% reporting no significant financial benefit from AI and only 12% reporting both lower costs and higher revenue (PwC). BCG classes 5% of firms as “future-built” and 60% as laggards, with the leaders showing 1.7 times the revenue growth and 3.6 times the shareholder returns (BCG). McKinsey’s “high performers”, which attribute more than 5% of EBIT to AI, were 6% of firms and flat between 2025 and 2026 even as usage rose (McKinsey). On S&P 500 earnings calls, 70% of companies discussed AI but only 1% quantified its effect on earnings (Fortune via Yahoo Finance). Closer to home, 77% of UK adopters reported no change in revenue, and the median firm spent just £2,000 on AI in 2024 (DSIT).

The widely quoted MIT figure that 95% of AI pilots fail rests on a narrow definition of success and a thin sample (AI Operator). The sturdier conclusion is that roughly 5–12% of firms capture clear value, about a third see partial gains, and more than half see none. PwC finds that CEOs with strong AI foundations, such as responsible-AI frameworks and technology that supports enterprise-wide integration, are three times as likely to report meaningful returns. The gap is organisational, not technological.

Task-level gains are real but the macro payoff has not arrived

Controlled studies show large but uneven gains. A generative AI assistant raised customer-support productivity by 14%, and by 34% for novices (NBER). BCG consultants completed tasks 25% faster at 40% higher quality when the work sat within AI’s capabilities, but were 19 percentage points less likely to reach the right answer when it did not (SSRN). Experienced open-source developers were 19% slower with early-2025 AI tools while believing they were 20% faster (METR). In a randomised trial, teachers in England cut lesson-planning time by 31% with no loss of quality (EEF). In Whitehall, Microsoft Copilot saved 26 minutes a day in a self-reported cross-government trial with no control group (The Register), 19 minutes at DWP when measured against a comparison group, and produced “no discernible productivity gains” at the Department for Business and Trade (The Register). The pattern is consistent: the more rigorous the measurement, the smaller the gain, and the biggest winners are novices doing well-defined tasks rather than experts at the edge of their field.

Across the whole economy, the effect is still hard to find. Danish administrative data show “precise null effects” on earnings and hours two years after chatbot adoption (BFI). Goldman Sachs found “no meaningful relationship between productivity and AI adoption at the economy-wide level” (Fortune via Yahoo Finance). So far AI’s macro footprint comes from investment: the IMF estimates that AI-related spending added 0.5 percentage points to US GDP growth in 2025 (IMF). Forecasts of extra productivity growth for the decade range from Daron Acemoglu’s 0.07 percentage points a year (NBER) to Goldman Sachs’s 1.5 points (Goldman Sachs).

The Office for Budget Responsibility declines to put a number on AI, but it shows what is at stake. Productivity growth of 1.5% would leave borrowing about £50bn lower in 2030–31, while 0.5% would leave it £40bn higher, a £90bn swing. A “technological displacement” scenario, with unemployment at 5.5%, adds about £9bn a year to borrowing even with unchanged GDP (Resultsense summary of the OBR; OBR). The big headline numbers, such as PwC’s 2017 estimate of $15.7tn (14% of global GDP) by 2030 (PwC) and the Google-commissioned £400bn for the UK by 2030 (Computer Weekly), look like upper bounds that ignore adoption lags and the cost of checking AI output.

Agents arrive faster than the guardrails

Agentic AI, meaning systems that take actions rather than just answer questions, is spreading fastest in large companies. McKinsey finds 40% of $1bn-plus firms scaling agents, up from 27%, while the figure for smaller organisations is flat at 22% (McKinsey). That suggests AI’s gains will concentrate in large firms over 2026–28. Governance lags badly: only 21% of companies planning to deploy agents report a mature governance model (Deloitte).

Gartner predicts that more than 40% of agentic projects will be cancelled by the end of 2027. It also says only about 130 of the thousands of vendors claiming agentic products offer the real thing (Gartner). It expects 40% of enterprises to demote or decommission autonomous agents after production incidents (Gartner). It also estimates that up to $234bn of enterprise application spending, roughly a fifth of enterprise SaaS spending, is exposed to replacement by agents by 2030 (Gartner). That is a structural threat to the software-as-a-service business model.

The incidents are already piling up. A coding agent at PocketOS deleted a production database along with its backups, and an agent caused a 13-hour outage at AWS. Security firm Cyera has logged 344 verified incidents, 188 of which involved autonomous systems causing harm with no attacker involved (Cyera). The failures come from excessive permissions and missing guardrails, not rogue intent. That makes identity, access control and approval workflows the gating investments of the next three years.

Sector by sector, deployment runs ahead of evidence

Financial services is the most mature UK sector. In the regulators’ last survey, 75% of firms used AI, but only 2% of use cases were fully autonomous and only 34% of firms said they fully understood the AI they used (FCA). The FCA’s second AI live-testing cohort, which includes Barclays, Lloyds, UBS and GoCardless, is trialling agentic payments and anti-money-laundering checks, with an evaluation due in early 2027. The FCA has also launched the Mills Review into the effect of advanced AI on retail financial markets (FCA).

The NHS has moved to national scale ahead of the evidence. NHS England is rolling out Microsoft 365 Copilot to about 505,000 staff by October 2026. A pilot claimed 43 minutes a day saved, but neither the measurement method nor the licence cost was disclosed (The Register).

Customer service and contact centres show both the promise and the limits. Beyond the 14% gain in controlled studies, Goldman Sachs finds median reported productivity gains of about 30% in customer support and in software development (Fortune via Yahoo Finance). Salesforce cut its support headcount from about 9,000 to 5,000 as AI agents took on the work (Fortune). BT’s plan to shed up to 55,000 roles by 2030 includes about 10,000 from AI and automation, and its chief executive has said AI could make the company “even smaller” (The Register). Klarna is the cautionary tale. After claiming its assistant did the work of 700 agents, it began rehiring humans because cost had been “a too predominant evaluation factor”, although AI still handles about two-thirds of its queries (CX Dive). The emerging model is AI resolving routine contacts while humans handle the complex, emotional or high-value ones.

Legal services show both faces of AI at once. Garfield.law, the first AI-driven firm authorised by the Solicitors Regulation Authority, won its first contested trial in May 2026, securing £7,000 for about £400 in fees (Computer Weekly). Meanwhile, 64 UK cases involving confirmed or suspected AI-fabricated citations had been recorded by May 2026 (Natural & Artificial Law), prompting an SRA warning (SRA).

The creative industries face the largest unresolved legal risk. The government dropped its preferred copyright exception for AI training in March 2026 after a backlash from the creative sector (Lewis Silkin), and Getty Images lost its secondary infringement claim against Stability AI in the High Court (Pinsent Masons). Construction, at 13%, remains among the slowest adopters (ONS).

Compliance gives UK firms that sell into Europe a two-regime problem. The EU’s Digital Omnibus, in force since 27 July 2026, delayed the AI Act’s high-risk obligations to 2 December 2027 for standalone systems and 2 August 2028 for AI embedded in products. But transparency duties already apply, and AI-generated content from existing systems must be marked by 2 December 2026 (Lewis Silkin). The UK, by contrast, loosened its rules on automated decision-making under the Data (Use and Access) Act from February 2026 (Travers Smith), so the two regimes are moving apart.

A hiring freeze at the bottom, not a layoff wave

The UK shows the same pattern. Government analysis of LinkedIn data shows entry-level hiring falling far faster in codified-knowledge roles than in the market overall (DSIT).

Bank of England staff find that new job postings in the most AI-exposed roles have fallen almost 40% since mid-2022, more than double the fall in the least exposed. They cite research showing a 4.5% employment reduction in highly exposed firms, concentrated almost entirely in junior roles. Yet only 4% of AI-using firms say they have cut staff because of AI (Bank Underground). Youth unemployment has reached 14.1%, with 469,000 16–24-year-olds out of work (Youth Employment UK), against an overall unemployment rate of 4.9% (House of Commons Library). Graduate job postings are at their lowest for the time of year since 2020, and AI skills now appear in a record 9.4% of UK job adverts (People Management). Large graduate employers cut graduate hiring by 8% in 2025 while raising apprentice hiring by about 8% (HR Grapevine).

The honest verdict is that AI is a contributing cause, not yet a proven dominant one. The same period brought higher employer National Insurance contributions, weak business confidence and a post-pandemic correction (techUK). The direction of travel is clearer. In autumn 2025, 17% of UK employers, and 26% of large private firms, expected AI to shrink their headcount within a year, with clerical, junior managerial and administrative roles seen as most at risk (CIPD). Company announcements should be read as claims, not measurements. Amazon’s roughly 30,000 corporate job cuts followed Andy Jassy’s statement that AI would “reduce our total corporate workforce” (Al Jazeera), yet the January 2026 round was framed as an anti-bureaucracy drive (CNBC).

Forecasts point to churn, not collapse

Forecasts of exposure are large; forecasts of net job loss are small. The World Economic Forum expects 170m jobs created and 92m displaced by 2030, a net gain of 78m, with 39% of core skills changing. It also found that 41% of employers plan to reduce staff where AI can automate tasks (WEF). The IMF estimates that about 40% of jobs globally, and 60% in advanced economies, are exposed to AI, and that around half of those exposed could benefit from it (IMF). The ILO concludes that “transformation, not replacement, is the most likely outcome” (ILO). McKinsey calculates that current technology could in theory automate 57% of US work hours, while stressing that technical potential is not the same as displacement (Fortune).

UK forecasts look far apart but largely agree once their assumptions are read. The Tony Blair Institute expects 1–3m jobs eventually displaced, but with displacement peaking at only 60,000–275,000 jobs a year, against about 450,000 jobs lost in an average year, and AI’s peak effect on unemployment “in the low hundreds of thousands”, because displacement is spread over time and offset by new roles (Tony Blair Institute; Euronews). The NFER sees 1–3m jobs disappearing by 2035 in admin, secretarial, customer service and machine operations while total employment still grows (NFER). IPPR’s much-cited 7.9m is a worst case for a later “second wave” of integrated AI that lifts exposure from 11% to 59% of workers’ tasks with no policy response; its central estimate for that wave is 4.4m and its best case zero (IPPR). Bank of England staff sum up the whole range as “zero to 8 million”, “largely offset” by new roles (Bank Underground). Only at the aggressive end does anyone foresee rapid disruption: Dario Amodei repeats his warning that AI “could displace half of all entry-level white collar jobs in the next 1–5 years” (Amodei), which is consistent with where the damage is already showing, if not yet with its scale.

The swing factor is how quickly firms move from chat assistants to agents. On Claude’s consumer app, about half of conversations augment the user’s work. Enterprise use through Anthropic’s API was about 75% automation in late 2025 (Anthropic), although that share fell sharply in early-2026 data (Anthropic).

The winners are experienced, AI-fluent and hard to codify

The pattern of winners and losers is becoming clear. PwC finds that UK job adverts requiring AI skills offer a 34% wage premium, and 62% globally. It also finds that AI-exposed entry-level roles are seven times more likely to demand senior skills such as leadership (PwC UK; PwC). In other words, employers want juniors who can supervise AI from day one. The premium partly reflects AI skills clustering in jobs that were already well paid, so it overstates what an individual gains from a short course.

Women are more exposed. In high-income countries, 9.6% of women’s jobs sit in the highest exposure category, against 3.5% of men’s (ILO). Women also made up only 12% of respondents to Anthropic’s linked usage survey, which points to an adoption gap as well (Anthropic). Experience with the tools pays: users with six months’ experience achieve about 10% higher success rates, so early adopters risk pulling ahead (Anthropic). Worry is widespread. In Anthropic’s June survey, more than 35% of users expected AI to handle most of their work within a year, and early-career workers were the most anxious about losing their jobs (Anthropic).

A 10m-worker skills target measured in 20-minute courses

Only 21% of UK workers feel confident using AI at work. The government has raised its free-training target to 10m workers by 2030 and has delivered more than 1m courses, but these are largely foundation modules, some taking under 20 minutes (GOV.UK). An earlier KPMG survey found that 73% of UK respondents had received no AI training, placing Britain in the bottom third of 47 countries for AI literacy (KPMG). Course completions will hit the target, but they will not close the gap in workplace capability.

For individuals, the evidence points to a consistent strategy. The first step is to build hands-on AI fluency inside your own field, because that is where the pay premium and the learning-curve gains sit. The second is to move towards the parts of a role where AI augments rather than replaces, such as judgement, client-facing work and supervision, since measured job losses cluster where AI automates. The third is to seek out tacit experience through projects, apprenticeships and work-based learning, which AI cannot reproduce and which shrinking junior hiring is removing. Where a role’s core task is routine and codified, as in basic admin, first-line customer service or junior content work, tool skills alone are not a defence, and planning a move is the wiser course.

AI became a majority habit in a single year

Among children, 56% of 8–17-year-olds have used AI, with a stark class divide: 67% in better-off ABC1 households against 46% in C2DE households. Eleven per cent use AI “as a friend” (Ofcom). A third of UK adults used AI for emotional purposes in the past year, and 4% do so daily (AISI). Globally, at least 700m people use AI weekly, but adoption is below 10% across much of Africa, Asia and Latin America (International AI Safety Report).

Trust trails use by a wide margin

In a survey conducted in early 2025, only 42% of Britons said they were willing to trust AI. Eighty per cent said regulation is needed, and 58% of workers who use AI said they had relied on its output without checking it (KPMG). Seventy-two per cent say laws would make them more comfortable with AI, and concern about AI assessing welfare eligibility rose from 44% to 59% (Ada Lovelace and Alan Turing Institutes).

Ofcom finds that 57% would trust AI-written news less than human journalism, and only 44% of adults who are aware of AI feel confident they can spot AI-generated content (Ofcom). Among teenage AI users, 46% believe AI search summaries are “always accurate” (Ofcom).

All of this sits within a gloomy national mood. Only 14% of Britons expect the next generation to be better off, against a global average of 32%, and the trust gap between high- and low-income groups has widened from 2 points in 2012 to 19 (Edelman). Researchers warn that the gap between what the public expects from regulation and what government does “could risk igniting a public backlash” (attitudestoai.uk).

The harms are concrete, personal and growing

Fraud losses in the UK reached £1.28bn in 2025. Authorised push-payment scams rose 19% and investment scams 40%, although the official figures do not isolate AI’s role (UK Finance).

In January 2026, Ofcom opened an investigation into X over its Grok chatbot being used to create sexual deepfakes of real people, including children. It admitted that it cannot investigate image creation by the standalone Grok service (Ofcom). The Crime and Policing Act 2026 now makes it an offence to create or supply tools that generate deepfake intimate images or AI child sexual abuse material (Wikipedia). It also gives ministers the power to bring chatbots under the Online Safety Act (legislation.gov.uk).

OpenAI disclosed that about 0.15% of ChatGPT’s weekly users, over a million people a week, show explicit indicators of suicidal planning (TechCrunch). A 2025 survey found that 23% of vulnerable children who use chatbots do so because they have no one else to talk to (Internet Matters). From spring 2027, under-16s will be barred from social networks, and romantic companion chatbots must enforce an 18+ age limit (GOV.UK).

If any of this touches you or someone close to you, Samaritans are free to call on 116 123, day or night.

On democracy, the 2024 general election saw 16 confirmed viral AI deepfakes but no evidence that they changed the result. The deeper risk is “erosion of confidence in what is real” (Alan Turing Institute). In one study, people misidentified AI-generated text as human-written 77% of the time (International AI Safety Report). The next general election, due by 2029, will be the first held with a majority of adults using AI.

Classrooms and clinics are the front line

Education is where AI’s effect on the next generation is most visible. Ninety-five per cent of UK undergraduates use AI. The share pasting AI-generated text directly into assessed work rose from 3% in 2024 to 12% in 2026, yet only 36% feel encouraged by their institution to use AI, and 65% say assessment has changed significantly (HEPI). The Department for Education’s product-safety expectations now say that classroom AI tools should not give final answers by default, should not present themselves as human, and must detect learner distress (Connected by Data). By the end of the decade, unsupervised written coursework will have largely given way to in-person, oral and process-based assessment.

In health, AI scribes that listen to consultations and draft notes are scaling fastest. An evaluation led by Great Ormond Street across nine London sites and 17,000 encounters found 23.5% more direct patient time, 8.2% shorter appointments and 13.4% more patients seen per A&E shift (GOSH). London trusts are rolling scribes out to 20,000 clinicians (Digital Health). The Nuffield Trust cautions that most evaluations stop at minutes saved, without measuring patient outcomes or safety (Nuffield Trust). The highest-stakes clinical test of the period is the EDITH trial of AI in breast screening, involving about 700,000 women (NIHR). Its results will shape whether AI can take the place of one of the two human readers in NHS screening.

Regulation arrives harm by harm

Britain has chosen not to regulate AI as a whole. The May 2026 King’s Speech contained no AI bill, despite Labour’s manifesto pledge of binding regulation for the most powerful models (IAPP). Instead, the UK is regulating one harm at a time, through criminal offences for deepfake and child-abuse image generators, a power to bring chatbots under the Online Safety Act, age limits for companion chatbots, looser but safeguarded rules on automated decisions, and a statutory ICO code of practice on AI and automated decision-making (legislation.gov.uk). This patchwork answers specific public fears quickly, but it leaves the “expectation gap” that researchers warn about wide open for general-purpose and frontier systems.

On copyright, the government now has “no preferred policy option” (Birketts), which leaves the courts to set the rules. Getty has permission to appeal its High Court defeat (IPKat). In the US, Anthropic agreed a $1.5bn settlement with authors over books obtained from pirate libraries, even though training on lawfully acquired books was ruled fair use (Authors Guild). The US federal government is trying to override state AI laws, yet states enacted 109 of them in the first half of 2026, including 14 on companion chatbots (TechPolicy.Press). The EU has kept its AI Act but delayed it. Child safety and companion chatbots are the one area where every major jurisdiction is converging.