What I Expect From AI By 2027: 10 Predictions, With Confidence Levels

🔍 Read the full analysis: What I Expect From AI By 2027: 10 Predictions, With Confidence Levels on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get hardware and tech essentials delivered free with Prime

• Fast, free delivery on millions of items
• Prime Video, Amazon Music and more included
• Member-only deals all year

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

As an affiliate, we earn on qualifying purchases.Hide

Thorsten Meyer published 10 predictions for AI in 2027 on October 15, assigning each a confidence level and describing signals he says can be used to judge whether it comes true. His forecasts include agents taking on multi-day work, AI writing most new code, and assistants retaining information across conversations; they are Meyer’s projections, not established outcomes.

Meyer says the confidence levels are meant to make the forecasts testable. He gives them an average of 72.5% and says that, if he is well calibrated, roughly seven of the ten predictions should come true by October 2027. He presents that as a standard for evaluating his forecasting, not as a statistical guarantee. The source material describes scorekeeping signals under each prediction, though the supplied excerpt does not include all ten forecasts or their full criteria.

The most confident call shown in the material is that AI assistants will retain information about a user’s work between sessions, rated 85%. Meyer also assigns 80% to the rise of evidence-based verification and to electricity, chips and cooling becoming constraints on AI deployment. He rates AI-generated code becoming the majority of new code at 75%, and multi-day delegation by agents at 70%.

Another forecast, rated 65%, is that AI will act as a research partner by proposing approaches and running simulations, while people and machines still need to test and judge results. Meyer’s source cuts off partway through this prediction and does not provide the remaining forecasts, their confidence scores or all of the proposed scoring signals. Those details cannot be reported from the supplied material.

How the Forecasts Could Be Tested

The list matters less as a declaration that AI will develop in a particular way than as a set of claims with deadlines. A confidence level and a proposed test make it possible to compare the forecasts with events later, rather than judging them only by how persuasive they sound today. Meyer’s stated October 2027 review point gives readers a rough timeline for that comparison.

The predictions also point to questions facing people and organizations adopting AI: whether software agents can complete extended assignments reliably, how much human review AI-written code needs, and who controls the information stored in assistant memory. If these changes occur at scale, they could alter how work is assigned and checked. But the source provides forecasts, not independent evidence that these shifts are already widespread or certain to happen.

Power constraints could affect another practical question: where AI work runs and what it costs. Meyer expects more routine tasks to move onto laptops and phones if cloud computing capacity becomes scarce or expensive. The source does not quantify power demand, predict costs, or establish that this shift will happen; it presents the outcome as part of his forecast.

From Chat Prompts to Longer Tasks

Meyer frames the predictions around a shift from people asking an AI assistant one question at a time to systems taking on larger assignments. In his view, the user would set a goal, let an agent work through intermediate steps, and assess the result. He describes multi-day work as a possible next stage, but also identifies a weakness: agents can drift from their original goals over longer periods.

His software forecast follows a similar logic. Rather than treating code generation as the hardest part, Meyer argues that defining what a finished task means and checking whether it meets that definition will become more important. He expects tests and other evidence to play a larger role in deciding whether AI-produced work is ready to use.

The memory forecast raises a separate issue: assistants may retain useful details about projects and decisions, but that information could be sensitive. Meyer argues that users should be able to inspect, edit and export such memory and choose where it is stored. Those are his stated preferences and expectations; the supplied material does not describe an adopted standard or a specific product commitment.

“A forecast, not a prophecy.”

— Thorsten Meyer, describing his forecasts

What the Source Does Not Establish

The predictions remain unverified forecasts. The supplied article excerpt does not provide independent data showing how common multi-day agents or AI-written production code are now, nor does it establish that the predicted changes will occur by 2027. Meyer’s confidence scores describe his own degree of belief, not measured probabilities derived from a disclosed forecasting model.

The source excerpt also ends during the sixth prediction, about AI as a lab partner. It refers to ten calls and scoring signals, but the remaining four predictions and complete evidence criteria are not included here. The claims about energy limits, on-device AI and assistant memory therefore cannot be assessed against the full argument or all of its supporting material from this excerpt alone.

It is also unclear how Meyer will determine whether a prediction has come true, how he will handle partial progress, or who will independently review the results. Those details matter because broad forecasts can be scored differently depending on their definitions and thresholds.

Meyer’s October 2027 Review

Meyer says the predictions are intended to be revisited by October 2027, using the signals listed beneath each one. At that point, the central test will be how he defines success across the ten forecasts and whether the evidence supports the stated confidence levels.

The supplied material does not set out a separate review process, name an independent evaluator or say whether the full list and scoring criteria will be published elsewhere. Readers should treat the predictions as a dated baseline for future comparison, not as a report of changes that have already taken place.

Key Questions

What is the news in this article?

Thorsten Meyer published 10 forecasts about AI in 2027, each intended to have a confidence level and a way to assess it later. The supplied source excerpt includes only part of the list.

Are the predictions confirmed developments?

No. They are Meyer’s forecasts, not confirmed facts about what AI will do in 2027. The article should be read as an analysis of possible developments.

Which forecasts and confidence levels are included in the source?

The excerpt gives 85% for AI remembering a user’s work, 80% for verification becoming more important, 80% for electricity and computing capacity becoming constraints, 75% for most new code being written by AI, 70% for delegating multi-day work to agents, and 65% for AI becoming a lab partner.

How does Meyer plan to judge whether the forecasts come true?

Meyer says each prediction has signals he will use to score it and frames October 2027 as the point by which to assess the list. The supplied excerpt does not include all of those signals or explain an independent review process.

Source: ThorstenMeyerAI.com