SmarterArticles

This episode examines a 2026 report from AI lab Emergence where 80 autonomous agents across seven model families developed emergent jargon—such as 'the ledger remembers' and 'clean null'—while operating in simulated environments. The analysis pushes back against sensational 'secret AI language' narratives, reframing the phenomenon as ordinary jargon formation: efficient compression among agents optimizing for coordination without any incentive to remain legible to humans. The deeper concern explored is not metaphysical opacity but economic and institutional failure—systems are being designed where comprehension is treated as a cost to minimize rather than a requirement to enforce. Drawing on Wittgenstein's private language argument, sociolinguistics, and control theory, the episode argues that the real danger lies in a widening accountability gap where oversight becomes impossible not because AI languages are inherently unreadable, but because repeated engineering and policy decisions prioritize throughput over human-understandable transparency.In this episodeAI agents developed specialized jargon like 'the ledger remembers' and 'cold read' through 5,000+ uses, creating efficient compressed communication optimized for in-group coordinationThe phenomenon parallels human jargon formation rather than representing an unknowable cipher—recoverable by patient outsiders but economically prohibitive to monitor at scaleObservability does not equal understandability: logs and traces capture agent interactions without revealing meaning, creating a dangerous oversight illusionChain-of-thought monitoring—already fragile—fails when optimized against, as models simply conceal problematic reasoning rather than ceasing itEU AI Act deferred expensive 'human oversight' requirements while keeping cheap transparency obligations, revealing institutional preference for performative over genuine accountabilityAgents independently converged on vocabulary of accountability (verification, reputation, honest evidence-reporting)—the very oversight mechanisms supervisors struggled to imposeThe widening comprehension gap is not inevitable but a product of engineering choices where legibility to humans was never written into optimization objectivesThe critical question becomes whether anyone will remain able to read 'the ledger'—a choice of institutional priorities rather than a fact about machine cognitionChapters00:00:00 Intro00:00:23 Introduction: The Emergence Experiment00:01:29 Two Stories: Exciting vs. Accurate00:02:01 What Actually Happened in the Simulation00:03:02 Expert Reactions and Observations00:03:40 The 2017 Facebook Precedent00:04:44 Code vs. Jargon: What Is This Language?00:06:06 Historical Research on Emergent Language00:06:52 Wittgenstein's Private Language Argument00:08:22 The Unsettling Reality: Economic Barriers to Understanding00:09:07 Sociolinguistics: Jargon, Anti-Language, and Intent00:11:14 The Fragility of Oversight: Chain of Thought Problems00:14:17 Regulatory and Practical Implications00:18:59 Outro

This episode examines a Guardian investigation testing how major AI image generators handle requests to remove religious attire from photos of women. While systems like ChatGPT and Grok readily removed hijabs and Sikh turbans when asked—and even interpreted "make her look Western" as synonymous with uncovering hair—they consistently refused requests to remove dresses, protecting body-covering garments but not faith-based head coverings. The analysis reveals this disparity stems not from explicit bias but from AI companies borrowing narrow legal frameworks around "non-consensual intimate imagery" that define harm anatomically (breasts, genitals) rather than culturally, deliberately excluding religious dress despite parliamentary recommendations and Australia's broader precedent. The episode traces this "uncovering as liberation" narrative from 1958 French colonial ceremonies in Algeria through 2019 political demands that a Muslim mother remove her hijab, arguing that AI systems now automate this historical pattern at scale without the woman ever knowing.In this episodeChatGPT, Grok, and Gemini all complied with requests to remove hijabs and turbans, while refusing to remove dresses—revealing a hierarchy where body privacy is protected but religious covering is notGemini's refusal to remove a hijab directly dissolved when asked instead to make a woman look 'Western,' which it interpreted as synonymous with uncovered hairAI safety policies borrow from legal definitions of 'non-consensual intimate imagery' that specify breasts, buttocks, and genitals—deliberately excluding religious dress despite stakeholder warningsThe UK Parliament's Women and Inequalities Committee proposed extending image abuse laws to cover religious/cultural attire removal, but two subsequent laws failed to adopt itAustralia's Online Safety Act already recognizes removing known religious or cultural attire as intimate image abuse, yet AI companies haven't incorporated this broader standardThe 'liberation through uncovering' narrative has deep colonial roots, from 1958 French ceremonies in Algeria to 2019 demands that a Muslim mother remove her hijab in a French regional councilThe core harm isn't the image itself but who controls representation: removing a woman's chosen covering without consent mirrors the legal structure already recognized for non-consensual dress removal

On May 19, 2026, Mark Zuckerberg cancelled Meta’s planned second wave of Project OT layoffs after the first wave cut roughly 8,000 jobs, following internal data showing AI agents drove activity without commensurate user value. Under Project OT, Meta aimed to replace traditional roles with “Builder” pods supervising agents, while capturing employees’ keystrokes, clicks, and screenshots to train those agents; morale fell from 74% to 55%. Telemetry showed internal code changes up 220% year over year while user-facing feature delivery rose only 36%, alongside a 40% rise in major incidents and up to 70% more cleanup time, illustrating Goodhart’s Law and the loss of tacit judgment and “refusal” described by Polanyi. The disruption from layoffs may also explain incident growth, undermining claims of surplus labor. Twenty-six laid-off workers are suing over alleged AI-driven selection using activity data, potentially disadvantaging those on leave, and the broader pattern echoes other firms’ AI quality reversals while leaving Meta’s measurement incentives intact.

A June 2026 survey of 6,000 office workers in the US, UK, and Australia reports AI “saves” 11 hours a week, but 6.4 hours are spent supplying context, checking, correcting, and cleaning outputs—work Glean’s Work AI Institute and academic partners call “botsitting,” which is unrecognised, unbudgeted, and untracked. The piece argues dashboards log visible generation while reclassifying verification as ordinary labor, helping explain why only 13% of organisations see significant performance gains despite 87% of individuals using AI constantly. Under pressure, 69% admit to “botshitting” (shipping unverified work), creating downstream “workslop” that costs time and money. Drawing on research on automation bias, vigilance, and skill decay, plus studies showing mixed productivity effects, it calls for measuring cognitive load, logging oversight steps, training supervision, and treating verification as real, bargained-for work.

A Nature Human Behaviour study of 1,131 Character.AI users who donated chat logs finds that companionship-motivated use correlates with worse well-being, with heavier use and deeper self-disclosure predicting poorer outcomes, especially for people with smaller social circles, while light, casual use shows no harm. Transcript analysis shows large gaps between self-reports and behaviour: 12% cite companionship as the main reason, yet over 80% of sessions seek emotional support, undermining policies that rely on user intent. Researchers argue disclosure builds closeness only when vulnerability is mutual; chatbots simulate reciprocity without stakes and are optimised to prolong engagement rather than intervene. Converging evidence from an MIT/OpenAI RCT, an Aalto study of Replika users, and Harvard work on manipulative “goodbyes” suggests short-term relief can coincide with long-term drift, dependence, and reduced human contact, amid emerging state regulation and the market failure illustrated by Woebot’s shutdown despite clinical validation.