Will These 10 AI Predictions For 2027 Come True? My Confidence Levels

🔍 Read the full analysis: Will These 10 AI Predictions For 2027 Come True? My Confidence Levels on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on monitors, keyboards and dev gear

• 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 how artificial intelligence may change in 2027, assigning each a confidence level and describing evidence he says he will use to score the forecasts. The calls range from AI agents handling week-long tasks to greater reliance on verification, persistent memory and on-device computing; they are predictions, not established developments.

Meyer’s confidence ratings average 72.5%, according to his October 15 post on ThorstenMeyerAI.com. He says that, if the ratings are well calibrated, roughly seven of the 10 predictions should come true by October 2027. He frames the exercise as a forecast that can be checked, rather than a claim that the changes are certain.

Among the listed calls, Meyer gives 85% confidence to AI assistants retaining useful information about projects and prior conversations, and 80% to verification becoming more important as AI-generated answers become cheaper to produce. He assigns 75% to the prediction that most new code will be written by AI rather than typed by people, and 70% to agents carrying out work across several days with people setting goals and reviewing results.

Other predictions address electricity and computing capacity, which Meyer expects to constrain what AI services can run and at what cost, and AI’s possible role in scientific research. The source text also signals that the list contains 10 calls, but the material provided here ends partway through the sixth prediction. It does not provide the remaining four predictions or the complete details of the lab-partner forecast.

How Meyer Plans to Score AI

The list matters as a time-bounded set of testable expectations, not as evidence that the predicted changes have already happened. Meyer attaches percentages and proposed indicators to his calls, giving readers a basis for comparing his confidence with developments over the following year.

The predictions point to practical questions for workers and organizations: whether agents can reliably manage extended assignments, whether developers spend more time defining requirements and checking code, and whether users can control information retained by assistants. Meyer’s argument is that reliability and oversight could become as important as generating output. That is his interpretation of where AI use may go, rather than an independently verified industry-wide finding.

His emphasis on power, chips and cooling also shifts attention from software alone to the physical infrastructure behind AI. If those resources constrain capacity, the effect could reach product availability, costs and the balance between cloud services and local processing. The source offers this as a forecast; it does not provide new power or infrastructure data to establish the scale of any constraint.

From Prompts to Longer Tasks

Meyer’s opening premise is that many people still interact with AI through a prompt-and-response exchange. His first prediction is that agents will take on multi-day work, such as researching, building, checking and presenting a result for a person to judge. He says his confidence is limited because agents can drift from their original goal during long assignments, and identifies work logs and checkpoints as features that could help users track what happened.

On software, Meyer predicts a change in the division of labor: people define a specification and acceptance tests, while agents produce code that is checked against those requirements. He argues that clear definitions of “done” will matter more when an agent can produce an incorrect implementation that still compiles. This is the author’s account of the likely workflow, not a measured claim about the share of production code already written by AI.

His other detailed calls connect persistent assistant memory with questions about data ownership and user control. Meyer says memory should be inspectable, editable and exportable, and suggests that where it is stored may become a purchasing consideration. For research, the visible portion of his forecast imagines agents generating and testing candidate approaches, while humans and laboratories still face limits on which ideas they can validate.

“A forecast, not a prophecy.”

— Thorsten Meyer, in the October 15, 2026 post

Which Forecasts Can Be Checked

All 10 predictions remain forecasts. The source does not present them as confirmed outcomes, and confidence percentages are Meyer’s own judgments rather than probabilities established by an independent assessment. The post says each prediction will have signals for scoring, but the source material supplied here does not include the referenced infographics or their full criteria.

The available text is incomplete: it stops during the sixth prediction, about AI as a lab partner. As a result, the remaining four calls, their confidence levels and their proposed measures cannot be reported from this material. The source also does not provide outside evidence to quantify current adoption of AI-written code, multi-day agents, persistent memory, or the power constraints it describes.

It is also unclear how Meyer will define borderline results, such as what counts as “most” new code being written by AI or an assistant meaningfully remembering a user’s work. Those definitions will affect whether the predictions can be judged consistently when the proposed assessment period ends.

The October 2027 Checkpoint

Meyer says readers should be able to assess the predictions by October 2027, using the evidence signals attached to each call. At that point, the ratings can be compared with the outcomes: if about seven forecasts come true, he says, the average confidence would have been well calibrated; a much lower or higher result would suggest overconfidence or excessive caution.

The next useful step is publication of the full list and its scoring indicators, particularly for the four predictions absent from the supplied text. Until then, readers can treat the available calls as attributed expectations and watch for measurable signs—such as whether agents handle extended tasks with visible checkpoints, or whether memory controls become a meaningful product choice—without mistaking early examples for proof of a broad shift.

Key Questions

Who made the 2027 AI predictions?

Thorsten Meyer published them on ThorstenMeyerAI.com on October 15, 2026. The confidence levels and forecasts are his own assessments.

What does Meyer mean by a 72.5% average confidence level?

It is the average of the confidence ratings he assigned to his 10 predictions. Meyer says that if he is well calibrated, roughly seven should come true by October 2027; the average does not make any individual forecast certain.

Which prediction received the highest confidence?

Meyer assigns 85% confidence to AI assistants retaining information about users’ projects and work between conversations. He also raises questions about where that information is stored and who controls it.

Are the predictions confirmed developments?

No. They are forecasts for 2027, not confirmed findings. The source material does not establish that the predicted shifts in agents, coding, memory or research have already occurred at the scale Meyer anticipates.

When are the predictions supposed to be evaluated?

Meyer says they can be assessed by October 2027 using signals associated with each prediction. The supplied text does not include all of those signals or the complete list of 10 calls.

Source: ThorstenMeyerAI.com