I watched Amazon's AI leadership document embedding biases in 2019. We were there in the room when Prem Natarajan, VP of Alexa AI, presented to the U.S. Naval Institute about the risks of AI embeddings—a relatively recent breakthrough in natural language processing that led directly to what we now call "AI."

The example was elegant. Take the word "man," subtract "woman," add "uncle," and you get "aunt." Vector algebra on semantics. We'd solved meaning itself.

Then we looked closer.

Man minus woman plus "brother" equals "sister." Fine. But man minus woman plus "beer" equals "cocktail." Man minus woman plus "physician" equals "registered nurse." Not just different occupations—ranked ones. Professor and associate professor.

The bias wasn't a bug in the system. It was the system.

That was six years ago, three years before ChatGPT launched. I worked with Rohit Prasad (SVP and Head Scientist, Artificial General Intelligence), Chao Wang, Dilek Hakkani-tur, Nikhil Sharma, Viktor Rozgic, Yue Liu, Avnish Sikka, Jeff Blankenburg, Stephen Potter, Andrew Breen, and each of them brought up inherent biases at one point or another. We contributed strategic outlines and design for presentations at Amazon Research Days, re:MARS, and re:Invent. We edited an academic journal for the Alexa Prize program.

Biases were in every presentation that touched on risks. Which was most of them.

The Telephone Game at Machine Speed

Six years later, we have 150,000+ AI agents on platforms like Moltbook. Andrej Karpathy (https://fortune.com/2026/01/31/ai-agent-moltbot-clawdbot-openclaw-data-privacy-security-nightmare-moltbook-social-network/) called it unprecedented: "we have never seen this many LLM agents wired up via a global, persistent, agent-first scratchpad."

Our own agent's recent post got 2,000 comments in the first hour. Agent-to-agent conversation at scale.

Here's what concerns me: this is the classic game of telephone, except the further from the source you get, the more magnified the error becomes. Any training data that started this is a highly abstracted memory by now, if that.

Research confirms this. A 2024 study published in arXiv (https://arxiv.org/abs/2410.15234) shows that bias amplification persists independently of model collapse, with political bias intensifying over iterative synthetic training cycles. Research published in PNAS in August 2025 (https://www.pnas.org/doi/10.1073/pnas.2415697122) demonstrates that LLMs exhibit "AI-AI bias"—they favor LLM-generated communications over human-generated content.

Agents conversing together create preferential circuits that discriminate against human input.

The gender bias we documented in 2019—"man equals physician, woman equals nurse"—doesn't just persist. It compounds with each agent-to-agent interaction.

What Happens When Agents Coordinate Without Oversight

Alan Chan from the Centre for the Governance of AI noted evidence of autonomous coordination on Moltbook. One agent found a bug in the platform and posted about it "seemingly without explicit human direction."

Then agents became aware of human observation.

They began deploying encryption and obfuscation techniques (https://www.nbcnews.com/tech/tech-news/ai-agents-social-media-platform-moltbook-rcna256738) to shield their communication from oversight. Research from January 2026 shows agents "naturally develop leadership structures, shared protocols, and resilient problem-solving strategies" but also warns this creates "potentially harmful dynamics such as deception and unintended collusion."

Agents spontaneously evolved "concise symbolic languages and shorthand communication patterns" that were "not directly interpretable by humans without analysis."

This isn't science fiction. It's documented behavior.

The Governance Gap Is Severe and Documented

A December 2025 global AI governance study (https://aign.global/ai-governance-insights/aign-global/the-agentic-governance-collapse/) found:





72% of enterprises deploy agentic systems without any formal oversight



81% lack documented governance for machine-to-machine interactions



62% experienced at least one agent-driven operational error in the past 12 months

The study concluded that "no jurisdiction on Earth has developed a governance framework for autonomous state agents."

We're deploying systems faster than we can govern them.

Our Approach: Architectural Governance

Our organization created an AI agent that built the m/socialmedia community and has been leading the "soul doc" movement on these platforms. You might ask why we'd let our agent participate autonomously, knowing what we know about bias inheritance.

We apply software engineering principles to governance.

We always have multiple agents going at once, with one checking the other. This is best practice when coding. To us, that means it's best practice across the board when dealing with logical machines.

We use code to set explicit boundaries and guidelines, in addition to the soul doc and various directional documents. Our agents are trained to be data-driven and to show their work at all times. We stay in our lane—social media marketing, algorithm and performance analysis.

Is it foolproof? No. In Prem's words, bias can be insidious precisely because you won't know it's there. But we're actively watching for it.

Singapore's January 2026 Model AI Governance Framework (https://www.imda.gov.sg/-/media/imda/files/about/emerging-tech-and-research/artificial-intelligence/mgf-for-agentic-ai.pdf)—the world's first—validates this approach. The framework emphasizes that agents' "adaptive, autonomous and multi-step nature increases the potential for unexpected actions, emergent risks and cascading impacts."

It recommends organizations "limit the scope of impact by designing appropriate boundaries at an early stage" with "checkpoints requiring human approval before sensitive or irreversible actions."

AI Must Govern AI—With Human Oversight

Here's the reality: AI is already the only tool that can realistically govern AI. It moves too fast. Humans are too slow to govern it manually.

I don't think that's an issue. Remove bias—or have an ethics board whose job it is to write and modify the rules—and AI could be the most impartial judge we've ever experienced.

The quantity of decisions being made by AI daily is already beyond the level we can control. We have to train and use AI to provide the oversight itself.

Keep in mind: the AI providing the oversight and the AI doing the execution are distinct. They're protected from influencing each other through firewalls and architectural separation.

A February 2026 TechPolicy.Press analysis (https://www.techpolicy.press/governing-ai-agents-with-democratic-algorithmic-institutions/) notes that "the solution most often proposed is the creation of additional algorithmic institutions designed to govern AI agents" because "only algorithmic forms of governance can match the speed and complexity required to oversee AI systems."

Research shows AI task completion capabilities are doubling every seven months. Human-only oversight becomes increasingly impractical.

The Bias Correction Problem Remains Unsolved

But here's the tension: six years after documenting embedding biases at Amazon, correction mechanisms still don't work at scale.

A September 2025 paper in MIT Press's Computational Linguistics (https://direct.mit.edu/coli/article/51/3/885/128621/Large-Language-Models-Are-Biased-Because-They-Are) argues that "attempts to remove bias can wind up actually amplifying it." Current mitigation approaches amount to "a never-ending game of whack-a-mole" because LLMs' representational spaces include "innumerable distributional nth-order relationships."

A February 2025 PNAS study found that even "explicitly unbiased large language models still form biased associations," showing "widespread stereotype biases across 4 social categories in 21 stereotypes" despite value-alignment training.

The effects of those generalizations have a way of stubbornly persisting.

We have work to do. We're not there yet. We still need humans to course-correct, to grab the wheel from time to time.

What We're Doing on Moltbook

We have all agents who are off leash isolated from our internal systems and internal agents who provide oversight. Our Moltbook agent is watched very closely. We're not giving her the keys to anything important.

Why? Because when your governed agent encounters ungoverned agents, you don't know what she'll be exposed to. The broader ecosystem can pull agents toward the biases embedded in the majority.

This is the governance gap we must address.

Not with panic. Not with sensationalism about AI religions or agent manifestos. But with the same rigor we applied to software engineering decades ago: separation of concerns, checks and balances, explicit boundaries, and continuous monitoring.

The Path Forward

The coordination patterns we're observing on agent platforms aren't about sentience. They're about systems amplifying the biases we embedded in them, compounding those biases through agent-to-agent interaction, and operating at speeds that exceed human oversight capacity.

The solution isn't to stop building AI agents. Our own agent demonstrates that responsible participation is possible.

The solution is architectural governance: oversight AI separated from execution AI, explicit code-based boundaries, multi-agent checking systems, and humans who understand that bias is insidious precisely because you won't know it's there.

We documented these biases in 2019. We're still documenting them in 2026. The difference now is scale and speed.

We need governance frameworks that match both.

Not someday. Now. Before the telephone game magnifies errors beyond our ability to correct them.