My 2027 Predictions For AI: What I Expect And How Sure I Am

🔍 Read the full analysis: My 2027 Predictions For AI: What I Expect And How Sure I Am on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

Thorsten Meyer published 10 predictions for AI in 2027 on October 15, giving each a confidence rating and describing signals he plans to use to judge whether it comes true, building on broader AI forecasts for 2026. His forecasts include agents handling multi-day tasks, AI producing most new code, assistants retaining work context and electricity becoming a constraint on AI services, themes also reflected in AI predictions for business. They are personal forecasts, not confirmed developments.

Meyer says his confidence ratings average 72.5%, which he interprets as a rough expectation that about seven of the 10 forecasts will prove accurate by October 2027 if he is well calibrated. He presents the ratings as a way to make the predictions testable, rather than as measured probabilities derived from a formal forecasting model. The source does not describe an external review of the scores.

Among the higher-rated calls, Meyer assigns 85% confidence to AI assistants retaining information about users’ projects and decisions across conversations. He gives 80% to a shift toward checking AI outputs with tests, sources or other evidence, and the same rating to electricity and computing capacity becoming important limits on AI deployment. His most confident labor-related prediction is that people will type less new production code, rated at 75%.

Other forecasts concern agents carrying out work across several days, AI systems serving as research partners, and routine assistant tasks moving onto phones and computers as data-center capacity becomes constrained, topics covered in a forecast of AI technologies. Meyer’s source says each prediction is accompanied by signals for scoring it, but the supplied text ends partway through the discussion of the research prediction and does not provide the full list of all 10 calls or their complete scoring criteria.

How Meyer Plans to Score AI

The article matters as a dated, testable set of expectations about how AI may change everyday work and the infrastructure behind it. Rather than treating AI progress as a single question, Meyer separates forecasts about task delegation, software development, verification, data retention and energy use. Readers can compare those claims with developments over the following year.

Several predictions point to changes in responsibility, not just capability. If agents take on longer projects or write more code, people may spend more time defining goals and checking results. If assistants retain work histories, questions about data control and portability become part of product choice. If power and chips constrain services, capacity could affect availability, price and which tasks run locally. These are implications of Meyer’s forecasts, not outcomes established by the article.

The confidence scores also provide a way to assess the forecaster’s calibration. A single prediction can be right by chance; a set of predictions with stated probabilities is more informative when reviewed together. Meyer says the average rating implies roughly seven accurate calls, but the number alone does not establish that the forecasts are well calibrated.

A Forecast With a Review Date

Meyer published the predictions on ThorstenMeyerAI.com on October 15, 2026, and frames October 2027 as the point by which readers can assess them. He argues that a prediction should include a number and a way to determine whether it came true. The article therefore pairs confidence ratings with proposed disconfirming signals under individual forecasts.

The supplied material includes examples of those tests. For multi-day agents, Meyer says a forecast failure would be tools remaining mostly chat interfaces, with long-running work still limited to demonstrations. For AI-generated code, he says evidence against his call would be production code continuing to be typed line by line, with AI largely confined to autocomplete. For persistent memory, he points to assistants continuing to forget context or users not relying on memory features.

The source also links to Meyer’s related writing on software development and energy demand, and to a short video presenting the predictions. Those references show how he connects the 2027 calls to earlier arguments, but they do not independently verify the forecasts.

“A forecast, not a prophecy.”

— Thorsten Meyer

Which Forecasts Will Hold Up

The predictions remain unverified expectations, and the supplied source does not provide independent evidence that the forecasted changes will occur in 2027. Meyer’s confidence ratings are his own estimates. The article does not explain a statistical method behind them or compare them with other forecasters’ probabilities.

The provided source material is incomplete: it cuts off during the discussion of the AI-lab-partner prediction, after the statement that people deciding what to test will matter. It does not contain the remaining predictions, their confidence levels or all of the promised scoring signals. The exact full set of 10 calls therefore cannot be reconstructed from this material alone.

Some tests may also need clearer definitions before a fair review. For example, “most” new code depends on how production code is counted, while the extent of persistent AI memory could vary by product and user. Meyer’s source does not specify how borderline cases or changes in product availability would be handled.

The October 2027 Checkpoint

Meyer says the predictions are intended to be scored by October 2027, using the signals listed beneath each one. That review would show how many calls came true and whether the average confidence rating matched the overall results. The supplied material does not give a more detailed schedule or a separate independent scoring process.

Until then, relevant developments include whether AI agents can sustain multi-day work with visible logs and checkpoints; whether teams rely on AI for production code rather than autocomplete; whether assistants offer memory users can inspect, edit and export; and whether power and computing constraints shape where routine AI tasks run. Those are areas Meyer identifies in his forecasts, not confirmed milestones.

Key Questions

What did Thorsten Meyer publish?

He published 10 predictions for AI in 2027, each intended to include a confidence rating and evidence that could be used to score it. The supplied source text contains only part of the list.

How confident is Meyer in the predictions overall?

Meyer says the 10 confidence ratings average 72.5%. He estimates that about seven predictions would come true by October 2027 if he is well calibrated; this is his interpretation, not an independently verified result.

What are some of his highest-rated forecasts?

He gives 85% confidence to AI assistants retaining work context across conversations. He rates the growing importance of output verification and electricity or computing constraints at 80% each.

Are these confirmed developments?

No. They are Meyer’s forecasts about what may happen in 2027. Whether they prove accurate remains to be seen.

When are the predictions meant to be reviewed?

Meyer frames October 2027 as the point for checking the predictions against the signals described in his article. The supplied material does not set out an independent review process.

Source: ThorstenMeyerAI.com