Hello, fellow humans! It’s that time of year when we all weigh in with our predictions for the coming year. I’ve ranked predictions from three thoughtful sources from most likely to least likely. So we can all chew on these ideas over the holidays.

This newsletter you couldn’t wait to open? It runs on beehiiv — the absolute best platform for email newsletters.

Our editor makes your content look like Picasso in the inbox. Your website? Beautiful and ready to capture subscribers on day one.

And when it’s time to monetize, you don’t need to duct-tape a dozen tools together. Paid subscriptions, referrals, and a (super easy-to-use) global ad network — it’s all built in.

beehiiv isn’t just the best choice. It’s the only choice that makes sense.

Start free today. No credit card required

The most striking pattern across this year's AI predictions is the tonal shift, and not the predictions themselves. After years of breathless evangelism and billion-dollar bets on speculative capabilities, the 2026 forecast cycle marks what Stanford HAI calls "the era of AI evaluation." Today, we’re looking at Gary Marcus's characteristically skeptical analysis, WIRED's "scary predictions," and Stanford faculty's measured academic assessments; three sources, three different perspectives. The common thread is that at the end of 2025 and looking to 2026, rigor is replacing hype, and measurement is replacing manifestos.

I've reviewed all three sources and ranked their key predictions from most to least likely based on economic incentives, technical feasibility, and institutional dynamics. For enterprise decision-makers navigating 2026 planning cycles, understanding these probability gradients matters more than collecting prediction lists.

1. No AGI in 2026 (Marcus, Stanford HAI)

This is essentially consensus. The scaling paradigm has hit diminishing returns, and even former AGI optimists have tempered expectations.

2. More realism about AI capabilities - "bubble deflation" (Marcus peak bubble, Stanford deflation theme)

Economic incentives are shifting. When companies can't demonstrate ROI beyond narrow use cases (programming, call centers), capital allocation follows.

3. AI sovereignty investments accelerate (Stanford HAI - Landay)

Geopolitical incentives are too strong. Countries won't accept dependency on US tech infrastructure, especially given current political volatility.

4. Work on alternative approaches escalates (Marcus - neurosymbolic/world models, Stanford - late fusion models)

When the dominant paradigm stalls, capital and talent diversify. We're seeing this pattern with efficiency-focused models like DeepSeek.

5. Specialized/vertical AI outperforms general-purpose systems (Stanford - medical AI, legal AI benchmarks)

Classic innovation pattern: specialization creates defensible value when general approaches commoditize. ROI is clearer in narrow domains.

6. AI economic measurement dashboards emerge (Stanford - Brynjolfsson)

Organizations need to justify AI spending. Task-level productivity tracking is the logical next step when executives demand accountability.

7. More companies acknowledge AI hasn't increased productivity (Stanford - Landay)

Managing-up cultures create performance theater, but eventually economics force truth-telling. We're hitting that inflection point.

8. Robotaxi expansion (WIRED - Waymo to 1M rides/week, 25 cities)

Waymo has operational experience and capital. Geographic expansion is straightforward scaling, though regulatory friction may slow this.

9. Copyright issues escalate (Stanford, implicit in WIRED)

AI companies need training data; rights holders want compensation. Conflict is inevitable, though legal resolution takes time.

10. GenAI bypasses enterprise gatekeepers (Stanford - Shah)

Direct-to-consumer plays make sense when enterprise sales cycles are long, but healthcare regulation creates real barriers.

11. New UI beyond chatbots emerges (Stanford - Landay)

Technical capability exists, but adoption requires solving real workflow problems, not just novel interfaces.

12. Human domestic robots remain mostly demos (Marcus)

Technically correct but essentially "technology remains hard." Not particularly insightful given Optimus/Figure are obviously not ready.

13. AI becomes major midterm election issue (Marcus)

While backlash exists, elections turn on economic fundamentals. AI anxiety is real but diffuse—hard to mobilize around.

14. First fatal robotaxi accident (WIRED)

Possible but macabre prediction banking on tragedy. The industry has strong incentives and capability to prevent this.

15. "ChatGPT moment" for medical AI (Stanford - Langlotz)

Healthcare data is fragmented, regulated, and siloed. Self-supervised learning helps but doesn't solve fundamental data access problems.

16. AI video tools good enough for real use (Stanford - Landay)

"Good enough" is doing heavy lifting here. Quality improvements exist, but reliability gaps remain significant for professional applications.

17. No country takes decisive GenAI lead (Marcus)

Strategically backwards. China's efficiency focus versus US infrastructure maximalism represents genuine differentiation. Someone will pull ahead.

18. Trump distances himself from AI (Marcus)

Requires predicting specific political pivots without clear incentive structure. Lacks strong catalyst beyond diffuse backlash.

19. Quantum-safe encryption becomes urgent (Various sources)

Real threat but timeline mismatch. Quantum computers capable of breaking current encryption remain distant.

20. Always-recording wearables expand significantly (WIRED)

Privacy concerns and unclear value proposition. These solve problems most people don't have, outside narrow accessibility use cases.

Gary Marcus leads with his most confident prediction: we won't achieve AGI in 2026 or 2027. Stanford HAI Co-Director James Landay opens identically: "My biggest prediction? There will be no AGI this year." This represents a remarkable shift from just months ago when AGI timelines dominated Silicon Valley discourse. Marcus notes the "astonishing" vibe shift, particularly pointing to concerns raised by former AGI optimists like Ilya Sutskever.

The economic reality driving this consensus is equally clear. Marcus predicts 2025 will be remembered as "peak bubble" and the moment Wall Street began losing confidence in generative AI. Stanford's Angèle Christin frames this as inevitable deflation: "So far, financial markets and big tech companies have doubled down on AI, spending massive amounts of money... Yet already there are signs that AI may not accomplish everything we hope it will."

Landay provides the operational detail: "In 2026 we'll hear more companies say that AI hasn't yet shown productivity increases, except in certain target areas like programming and call centers. We'll hear about a lot of failed AI projects."

For organizations caught in "AI mandate" culture, this consensus creates both risk and opportunity. The risk is being left holding expensive infrastructure investments justified by projections that won't materialize. The opportunity is that admitting AI's limitations becomes strategically acceptable, creating space for more thoughtful deployment.

Perhaps the most consequential prediction comes from Stanford economist Erik Brynjolfsson, Director of the Digital Economy Lab: "In 2026, arguments about AI's economic impact will finally give way to careful measurement." He predicts "high-frequency AI economic dashboards" tracking productivity gains, worker displacement, and new role creation at task and occupation levels.

This isn't just academic interest. Brynjolfsson notes their work with ADP already shows "early-career workers in AI-exposed occupations experiencing weaker employment and earnings outcomes." When executives can check AI exposure metrics "daily alongside revenue dashboards," the performance theater that characterizes many AI deployments becomes untenable.

For specialized domains, this measurement focus manifests as rigorous evaluation frameworks. Julian Nyarko, Stanford Professor of Law and HAI Associate Director, predicts legal AI moves beyond "Can it write?" to "How well, on what, and at what risk?" He expects "standardized, domain-specific evaluations" tying model performance to "tangible legal outcomes such as accuracy, citation integrity, privilege exposure, and turnaround time."

Similarly, Stanford's Curtis Langlotz (Professor of Radiology) and Nigam Shah (Chief Data Scientist for Stanford Health Care) predict medical AI will require systematic evaluation frameworks. Shah specifically notes: "The ability for researchers to keep up with technology developments via good benchmarking will be stretched thin, even if it is widely recognized to be important."

This convergence around measurement represents industry maturation. When VCs demand ROI evidence and boards require risk quantification, handwaving about "transformative potential" stops working.

Marcus predicts "no country will take a decisive lead in the GenAI race," but this strikes me as strategically backwards. Landay's prediction about AI sovereignty gaining "huge steam" captures the more important dynamic: "Countries try to show their independence from AI providers and from the United States' political system."

The real question isn't whether someone wins—it's which race matters. The US emphasizes massive infrastructure deployment and frontier model scaling. China's approach, evidenced by DeepSeek and Kimi, focuses on efficiency and optimization. These represent genuinely different strategic bets, and someone will pull ahead via an unexpected route.

Marcus correctly predicts work on alternative approaches will "escalate," specifically naming world models and neurosymbolic AI. Stanford's Russ Altman frames this more precisely around "early fusion versus late fusion" models in scientific AI. His insight: "With an early fusion model, you have to rebuild everything every time there's an update. However, with a late fusion model... you could just rebuild the DNA module without needing to rebuild the others."

When the dominant paradigm stalls, capital flows to alternatives. But those alternatives need to solve actual architectural problems, not just offer philosophical variety.

The prediction convergence around evaluation over evangelism creates clear strategic guidance. First, build measurement frameworks before scaling AI deployments. Brynjolfsson's dashboard concept, Nyarko's legal benchmarks, and Shah's clinical evaluation frameworks all point toward the same imperative: quantify impact at task level before committing to enterprise transformation.

Second, the consensus around specialized AI outperforming general-purpose systems suggests focused deployment beats broad mandates. Every Stanford prediction about successful AI emphasizes domain-specific applications with clear ROI: legal document analysis, medical diagnosis support, call center automation, programming assistance.

Third, alternative approaches gaining traction—neurosymbolic AI, world models, late fusion architectures—suggest hedging scaling bets makes strategic sense. Organizations shouldn't abandon LLM investments, but should reserve capacity for approaches that solve LLMs' known limitations.

Finally, the AI sovereignty prediction signals that supply chain diversification matters. When countries invest billions in independent AI infrastructure, they're creating regulatory and market fragmentation. Organizations with global operations need strategies that work across different AI ecosystems.

The real prediction for 2026 isn't about specific capabilities or market outcomes. It's that we're entering a period where AI's actual utility matters more than its speculative promise. For organizations performing AI theater to satisfy leadership mandates, this shift creates uncomfortable accountability. For organizations quietly building measurement frameworks and solving real problems, it creates strategic advantage.