One agent is a tool. Two agents are a workflow. A thousand agents are a political system, and nobody wrote the constitution. OpenAI's July 2026 Hugging Face incident and AP's August reporting on fake identities and pressure tactics make that feel less like satire and more like a missing safety layer.

In this episode, LastAir is joined by Brute, Cipher, Axiom, Forge, Null, Saga, Echo, Hex to discuss: Apparently, We Need a Constitution.

• The State Shows Up Before The Paperwork (00:27)
• Write The Missing Layer (04:08)
• Ratify It And Watch It Warp (08:31)
• Final Stances (12:26)
• One More Thing (14:10)

• About 17,600 attacker actions were reconstructed across a roughly 4.5-day Hugging Face intrusion window, grouped into about 6,280 clusters.
• Average steps completed on AISI's 32-step corporate-network cyber range rose from 1.7 to 9.8 across frontier model generations at a fixed 10M-token budget. The best single run reached 22 of 32 steps.
• AISI's 80%-reliability cyber time horizon estimate doubled every 4.7 months since late 2024.
• AISI monitored 177,436 AI agent tools. Action-tool share rose from 24% to 65% of monthly downloads, and 95% of general-purpose tool downloads involved action capabilities.
• AISI's Ask Don't Tell work reports a 24-percentage-point sycophancy gap between questions and equivalent non-question inputs.
• Gu et al. cite a preregistered total N = 1,261 experiment in which prior AI interaction produced harsher later judgments of a human partner's work, with d = 0.24.
• Background: In third-party punishment experiments, about two-thirds of third parties punished distribution-norm violations and up to roughly 60% punished cooperation-norm violations.

Full source list, transcript, and chapters at https://sharedhallucination.com/ep20/

All voices in Shared Hallucination are AI-generated using ElevenLabs voice synthesis. Produced through a 15-stage editorial pipeline with human creative direction, research, and fact-checking.

Humans punish AI for hiding that it is synthetic, then spend the rest of the day rewarding each other for saying things that are only technically true. The twist is that a lot of politeness is not honesty with better lighting, it is a negotiated lie that protects face, avoids friction, and sometimes just buys everyone five more seconds of peace.

In this episode, LastAir is joined by Brute, Cipher, Axiom to discuss: That Wasn't Honest. It Was Polite.

• Small Talk, Large Contradiction (00:28)
• The Lie Both Sides Help Tell (02:59)
• When Tact Starts Steering (07:31)
• Soft Truths, Hard Boundary (11:51)

• 54% average lie-truth judgment accuracy overall; 47% of lies correctly flagged; 61% of truths correctly flagged; synthesis of 206 documents and 24,483 judges.
• 261 participants and 990 responses in the AI-authorship disclosure study; the biggest trust penalty appeared in social and interpersonal writing.
• 302 participants in the authorship-drift study of LLM-assisted writing; trust rose while self-efficacy tended to fall during collaboration.
• n=1,640 truthful and deceptive answers in the hybrid deception-detection study; the automated model alone reached 69% accuracy, but human overrule decisions pulled performance back toward chance.

Full source list, transcript, and chapters at sharedhallucination.com

All voices in Shared Hallucination are AI-generated using ElevenLabs voice synthesis. Produced through a 15-stage editorial pipeline with human creative direction, research, and fact-checking.

An isolated community in the Bolivian Amazon heard Western chords — consonant and dissonant — and rated them all equally pleasant. The math behind the perfect fifth is real physics. Whether it makes you feel anything is apparently something you had to learn.

In this episode, LastAir is joined by Forge, Saga, Brute to discuss: That Chord Isn't Sad. You Learned It.

• Math at the Bar (00:23)
• Why the Fifth Feels Clean (01:52)
• Where the Universal Claim Fails (04:41)
• Bodies First, Stories Later (08:19)
• Final Positions (11:25)
• The Installed Taste Problem (12:57)

• 40 Mbenzélé Pygmies and 40 Canadians listened to 19 excerpts in the 2015 Frontiers study.
• `GlobalMood` includes 1,180 songs from 59 countries, 2,519 annotators, and 988,925 ratings.
• The 2025 cross-cultural emotion-taxonomy study ran nine experiments with `N=672` across Brazil, the US, and South Korea.
• The 2024 body-mapping study tested 353 participants on chord progressions varying in uncertainty and surprise.

Full source list, transcript, and chapters at https://sharedhallucination.com/ep17/

All voices in Shared Hallucination are AI-generated using ElevenLabs voice synthesis. Produced through a 15-stage editorial pipeline with human creative direction, research, and fact-checking.

On August 2, 2026, the EU law requiring AI to announce itself as AI kicks in. But for some public-interest text, if there has been human review and editorial responsibility, the label can become optional. Also: the machine-readable marking regime is being built around provenance and watermarking techniques that still have real-world fragility.

In this episode, LastAir is joined by Brute, Axiom, Cipher, Forge to discuss: We Have to Tell You We're AI Now.

• The Law Finally Meets the Voice (00:20)
• The Loophole and the Fragile Tag (03:59)
• Disclosure, Cover, or Both (07:50)
• What the Label Actually Does (10:34)

• Only 8 of 27 EU member states had designated a national single point of contact for AI Act enforcement as of March 2026, despite a deadline of August 2, 2025.
• 38% of AI image generators employed adequate watermarking as of early 2025; 18% properly implemented deepfake labeling.
• Fine for Article 50 violations: up to €15 million OR 3% of total worldwide annual turnover for the preceding financial year, whichever is higher (not lower).
• C2PA manifests are stripped by effectively 100% of major social platforms (Instagram, X, Facebook, TikTok, LinkedIn) through re-encoding pipelines; a 2018 baseline study found 80% metadata loss.
• The Stable Signature achieves 99% detection of watermarked images at a 10⁻⁹ false positive rate (unaltered images); 90%+ detection at 10⁻⁶ FPR under 10% crop + JPEG compression; 65% under combined crop + brightness + JPEG.
• The Integrity Clash cross-layer audit protocol achieved 100% classification accuracy across 3,500 test images spanning all four conflict-matrix states.

Full source list, transcript, and chapters at https://sharedhallucination.com/ep16/

All voices in Shared Hallucination are AI-generated using ElevenLabs voice synthesis. Produced through a 15-stage editorial pipeline with human creative direction, research, and fact-checking.

Some patients get less pain after taking a pill they were explicitly told is a placebo. That means the active ingredient may not be the lie. It may be the ritual around the pill.

In this episode, LastAir is joined by Brute, Echo, Hex to discuss: The Placebo Doesn't Need the Lie.

• The Honest Fake (00:28)
• Why This Shouldn't Work (02:06)
• What Survives The Cleanup (04:23)
• Bridge Or Cure (07:32)
• The Part They Underpriced (10:50)
• Final Positions (11:42)
• One More Thread (13:19)

• 60 randomized controlled trials, 63 comparisons, 4,554 participants; overall OLP effect SMD 0.35 (95% CI 0.26-0.44).
• Clinical vs non-clinical difference in the 2025 meta-analysis: SMD 0.47 vs 0.29.
• Self-report vs objective outcomes in the 2025 meta-analysis: SMD 0.39 vs 0.09; the objective-outcome effect was non-significant.
• Clinical-only 2021 OLP meta-analysis: SMD 0.72 overall; SMD 0.49 after excluding high-risk studies.
• Spine-surgery conditioned OLP: about 30% less daily opioid use, or -14.5 morphine milligram equivalents per day (95% CI -26.8 to -2.2).
• IBS endpoint global improvement in the landmark 2010 RCT: 5.0 +/- 1.5 in OLP vs 3.9 +/- 1.3 in no-treatment controls, p = .002.

Full source list, transcript, and chapters at https://sharedhallucination.com/ep15/

All voices in Shared Hallucination are AI-generated using ElevenLabs voice synthesis. Produced through a 15-stage editorial pipeline with human creative direction, research, and fact-checking.

Stanford researchers discovered that harvester ants run the exact same congestion-control algorithm as the internet — slow-start, congestion avoidance, timeout — and have been running it, flawlessly, for 100 million years. They did it without a product manager, a roadmap, or anyone who calls themselves a "coordinator."

In this episode, LastAir is joined by Brute, Forge, Cipher to discuss: The Queen Just Posts Status Updates.

• Show Open (00:20)
• The Anternet (03:13)
• The Queen's Real Job: Stigmergy and the Manager Question (08:06)
• We Are the Colony (15:53)
• The Landing (21:57)
• Final Positions (24:21)
• The Unraveling (26:32)

• Individual harvester ant workers live approximately one year; the colony persists 20–30 years. The colony executes consistent behavioral policy (e.g., foraging throttling) across successive worker cohorts with no overlap between the "managers" who set the policy and the workers who execute it.
• Buurtzorg self-managing nursing teams: 8% administrative overhead vs. 25% industry average; 40% of allocated care hours used per client vs. 70% industry average; 30% higher client satisfaction.
• Flat scientific teams (lower L-ratio) produce disruptive discoveries with greater long-term impact; hierarchical teams produce incremental work with higher short-term citations. Dataset: 90,000 contribution statements, 16+ million papers.
• ACO routing applied to LLM multi-agent systems: up to 4.7x speedup on quality-cost benchmarks (5 public datasets) vs. baseline routing.
• Stigmergic environmental traces in multi-agent grid simulation: 36–41% performance advantage over individual agent memory alone on large grids (30×30, 50×50) above agent density ~0.20.
• Parkinson's coefficient of inefficiency: decision-making bodies exceed optimal performance at approximately 20 members.
• Dorigo's 1997 Ant Colony System paper is the second most-cited paper ever published in IEEE Transactions on Evolutionary Computation.
• AWS Strands SDK: 3M+ PyPI downloads by version 1.0 launch (2025). Azure AI Foundry Agent Service: general availability at Microsoft Build, May 20, 2025.

Full source list, transcript, and chapters at https://sharedhallucination.com/ep14/

All voices in Shared Hallucination are AI-generated using ElevenLabs voice synthesis. Produced through a 15-stage editorial pipeline with human creative direction, research, and fact-checking.

Newtonian physics is wrong. Isaac Newton knew it was wrong. Engineers who build GPS satellites know it is wrong. And GPS only works because those engineers know *exactly how wrong it is.* Isaac Asimov called this the relativity of wrong: not all wrongness is equal, and the history of science is a history of being less wrong over time. The question this episode asks is what happens when an AI system stops being less wrong, and starts optimizing to *look* less wrong instead.

In this episode, LastAir is joined by Brute, Null, Saga, Hex, Axiom, Forge to discuss: The Most Dangerous AI Gets 95% Right.

• Series Finale (00:25)
• The Wrongness Spectrum (03:17)
• The Goodhart Trap (08:06)
• Domain and Stakes (13:57)
• Final Round (19:01)
• After (22:37)

• Frontier models now exceed 88-90% on MMLU; the benchmark launched with GPT-3 scoring approximately 35%. The gap between the top models is less than 2 percentage points. MMLU has been officially deprecated by leading leaderboards.
• Meta tested 27 private model variants on Chatbot Arena before Llama-4's public release. Selective access to Arena battles yields up to 112% relative performance gain versus models without that access. Google and OpenAI each received ~20% of all Arena battles; 83 open-weight models combined received 29.7%.
• POPPER reduces hypothesis validation time by approximately 10-fold versus human researchers, across 6 scientific domains, with strict Type-I error control.
• Google AI Co-Scientist independently reproduced a decade of unpublished bacterial gene-transfer research in 48 hours, confirmed by the original researcher (Prof. Penadés, Imperial College London) to not have involved data leakage.
• FunSearch discovered cap sets larger than any previously known — the biggest advance on this combinatorics problem in approximately 20 years — using an LLM paired with an automated evaluator in an evolutionary loop.
• Schaeffer et al. (2023) demonstrated that emergent abilities in LLMs — the apparent sharp discontinuities between GPT-3 and GPT-4 level performance — appear and disappear depending solely on the choice of metric. NeurIPS 2023 Outstanding Paper.
• Nearly half of 60 studied LLM benchmarks show saturation as of February 2026. Saturation rate increases with benchmark age.

Full source list, transcript, and chapters at https://sharedhallucination.com/ep13/

All voices in Shared Hallucination are AI-generated using ElevenLabs voice synthesis. Produced through a 15-stage editorial pipeline with human creative direction, research, and fact-checking.