ResearchPapers, evidence & method
Gabriel Unger and Arvind Narayanan Propose a Research Agenda to Steer AI From Full Automation Toward Human Collaboration
Gabriel Unger and Arvind Narayanan argue that AI research is overly fixated on full automation. They say a system can achieve only two of autonomy, reliability and generality, and they set out four technical problems for building AI that works with humans.
From the AI as Normal Technology article of October 14, 2026 and cited sources: The central claim is a trilemma: an AI system can be fully autonomous, reliable in high-stakes settings or broadly general, but only two of these at once. [1] [2] [8] [9]
SOURCES[1] Steering AI Beyond Automation (normaltech.ai) normaltech.ai[2] The Turing Trap (arXiv) arxiv.org[8] Acemoglu & Restrepo, AI, Automation and Work (TSE) tse-fr.eu
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- The central claim is a trilemma: an AI system can be fully autonomous, reliable in high-stakes settings or broadly general, but only two of these at once. [1] [2] [8] [9] [10]
- Original source
- AI as Normal Technology article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- Studies
- arXiv 2201.04200 (Jan 2022) [2] date from arXiv ID
- When
- October 14, 2026 article published
- Covered
- Oct 15 edition
- Evidence
- 11 sources across 11 sites
- 1Steering AI Beyond Automation (normaltech.ai)normaltech.ai
- 2The Turing Trap (arXiv)arxiv.org
- 3The Decoder: AI coding can make developers slowerthe-decoder.com
- 4eWeek: Developers using AI take 19% longereweek.com
- 5TechCrunch: Meta will record employees' keystrokestechcrunch.com
- 6SiliconANGLE: Thinking Machines interaction modelsiliconangle.com
BreakoutHow it’s measured
Substack · 119 likes · 10 comments
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
The central claim is a trilemma: an AI system can be fully autonomous, reliable in high-stakes settings or broadly general, but only two of these at once. The authors present this as a reason to prioritize human–AI collaboration. Their view builds on the Turing Trap argument and on Acemoglu and Restrepo's work suggesting that automation can be adopted faster than is efficient, at the expense of technologies that raise productivity in other ways. [1] [2] [8] [9] [10]
The agenda names four technical problems. First, agents should show their sources, assumptions and uncertainty as they work, which is hard outside coding because few fields keep records of their work processes. Second, low latency for collaboration requires changes across inference, retrieval, tools and browsers. Third, agents need shared workspaces that allow interruption instead of strict turn-taking. Fourth, evaluation is needed where simulated users may not match real ones and can reward the wrong behavior. [1]
Two recent industry moves relate to the agenda. Meta now records employees' clicks and keystrokes to train agents, which has drawn privacy and labor objections. Thinking Machines Lab released TML-Interaction-Small, a full-duplex model, to selected partners only. Its company-reported turn-taking time of under 0.4 seconds has not been independently reproduced. [5] [11] [6] [7]
the agenda is argument, not experiment, though METR's trial and Meta's data capture show related dynamics.
Read the full assessment
Implication: builders and buyers may need to measure joint human–system performance, expose intermediate steps and design for interruption, especially in fields that require a human decision-maker.
Executive brief
Gabriel Unger, a Stanford Digital Economy Lab economist, and Princeton's Arvind Narayanan argue that AI's main research program is far too focused on full automation. They propose four technical problems that would move it toward human–AI collaboration (Steering AI Beyond Automation). Their most pointed claim is a trilemma: an AI system can be fully autonomous, reliable in high-stakes settings or truly general-purpose, but only two of the three at once. The essay is commentary. It offers no new experiments. It does connect to things that have actually happened: Meta now records its employees' keystrokes to train agents, and Thinking Machines Lab has released a research preview of a low-latency "interaction model."
What changed and event timeline
The "Turing Trap" framing
Brynjolfsson argues that technologists, executives and policymakers face too strong an incentive to automate rather than augment, and that this concentrates power. The new essay cites this argument directly ().
METR coding RCT
In a randomized trial, 16 experienced open-source developers were 19% slower on tasks when AI tools were allowed, though they believed the tools had made them about 20% faster (;).
Meta captures employee inputs
Meta began recording staff mouse movements, clicks and keystrokes in selected apps to train agents. The company says safeguards protect sensitive content ().
Thinking Machines interaction models
Thinking Machines Lab announced TML-Interaction-Small, a full-duplex model that processes input and output in 200 ms chunks. It is available only as a research preview for selected partners ().
Collaboration research agenda published
Unger and Narayanan lay out four collaboration problems and a bottom-up, market-driven way to steer AI away from pure automation ().
Capabilities and access
The essay is a research agenda, so there is no model to access. The closest shipped example it points to is Thinking Machines' TML-Interaction-Small. Access is limited to selected research partners, with a wider release promised for later in 2026.
Read the full section
The essay is a research agenda, so there is no model to access. The closest shipped example it points to is Thinking Machines' TML-Interaction-Small. It is a mixture-of-experts model with 276B total parameters, about 12B of them active at a time. A fast "interaction model" handles the conversation and hands heavier reasoning and tool calls to an asynchronous "background model" (SiliconANGLE; MarkTechPost). Access is limited to selected research partners, with a wider release promised for later in 2026.
Technical analysis for researchers and developers
- Agents that expose their sources, assumptions, methods and uncertainties as they work.
- Getting there means changing inference, retrieval, tools and browsers, not just the model.
- Agents need shared workspaces, live state and tolerance for interruption and simultaneous editing by a human.
Read the full section
The essay's four problems (normaltech.ai):
- White-box agents. Agents that expose their sources, assumptions, methods and uncertainties as they work. The main obstacle is data: coding has public repositories that record every step of development, but law, medicine and many sciences have no comparable process records.
- Collaboration-grade latency. Getting there means changing inference, retrieval, tools and browsers, not just the model.
- Beyond turn-taking. Agents need shared workspaces, live state and tolerance for interruption and simultaneous editing by a human.
- Collaboration evaluation. Fixed benchmarks can't capture work that involves a human. Simulated users help, but small gaps between simulated and real users can push training to reward the wrong behavior.
Claims and evidence
- Coding agents erode "flow" and cause burnout. The essay bases this on programmer reports and does not quantify it (normaltech.ai).
- Firms are racing to capture "process traces." This is confirmed for Meta (TechCrunch).
- Low latency is achievable.
Read the full section
- Coding agents erode "flow" and cause burnout. The essay bases this on programmer reports and does not quantify it (normaltech.ai). METR's trial is related evidence: it found developers slowed down while believing they had sped up (eWeek).
- Firms are racing to capture "process traces." This is confirmed for Meta (TechCrunch). The essay's claim about Mercor has no independent source among the reviewed sources.
- Low latency is achievable. The FD-bench turn-taking figures (under 0.4 s for TML-Interaction-Small, 0.57 s for Gemini-3.1-flash-live, 1.18 s for GPT-realtime-2.0) are reported by the company and have not been reproduced independently (SiliconANGLE).
Context and prior work
- The essay traces its view to Licklider's 1960 "Man-Computer Symbiosis" and Engelbart's 1962 "Augmenting Human Intellect" (normaltech.ai).
- Economics: Acemoglu and Restrepo's task-based model suggests automation may be adopted faster than is efficient, at the expense of other productivity-raising technologies (TSE working paper; AER 2018).
- The essay says these economists remain a minority in their field and that many have dismissed the Acemoglu–Johnson argument for pro-augmentation AI.
Read the full section
- The essay traces its view to Licklider's 1960 "Man-Computer Symbiosis" and Engelbart's 1962 "Augmenting Human Intellect" (normaltech.ai).
- Economics: Acemoglu and Restrepo's task-based model suggests automation may be adopted faster than is efficient, at the expense of other productivity-raising technologies (TSE working paper; AER 2018). Brynjolfsson argues that augmentation creates more value than imitating humans (Daedalus).
- The essay says these economists remain a minority in their field and that many have dismissed the Acemoglu–Johnson argument for pro-augmentation AI.
Limitations, safety and contested findings
- The trilemma has no data behind it. The autonomy/reliability/generality trade-off is argued, not measured.
- METR's trial is narrow. It covered 16 developers using early-2025 tools and is an imperfect stand-in for today's agents (The Decoder).
- Process-trace capture raises labor and privacy concerns. The essay notes that this data mostly feeds black-box automation today.
Read the full section
- The trilemma has no data behind it. The autonomy/reliability/generality trade-off is argued, not measured.
- METR's trial is narrow. It covered 16 developers using early-2025 tools and is an imperfect stand-in for today's agents (The Decoder).
- Process-trace capture raises labor and privacy concerns. Meta says it has safeguards, but critics raise privacy and labor-rights objections (OECD AI incident). The essay notes that this data mostly feeds black-box automation today.
- Collaboration lacks agreed metrics. The authors acknowledge that standard benchmarking may not work for collaboration at all.
Business and practitioner implications
- The essay proposes research prizes, tax changes that stop favoring automation, government purchasing power, oversight requirements stronger than a rubber stamp, competition rules against vendor lock-in, and a voice for workers in how AI is deployed (normaltech.ai).
- Regulated fields that legally require a human decision-maker are the authors' bet; they name Harvey (law) and OpenEvidence (medicine).
- For builders:
Read the full section
- Policy levers. The essay proposes research prizes, tax changes that stop favoring automation, government purchasing power, oversight requirements stronger than a rubber stamp, competition rules against vendor lock-in, and a voice for workers in how AI is deployed (normaltech.ai).
- Where it fits first. Regulated fields that legally require a human decision-maker are the authors' bet; they name Harvey (law) and OpenEvidence (medicine).
- For builders:
- Measure how humans and the system perform together, not only autonomous task completion.
- Show intermediate steps to users.
- Design for interruption.
- Treat proprietary process-trace data as a strategic asset.