AI agents and synthetic activity

Social signals in crypto rest on a premise that used to be safe: a post indicates a person, and a lot of posts indicate a lot of people paying attention. Text generation has made producing plausible, varied, on-topic posts at scale essentially free, and the premise no longer holds automatically.

This page is about the measurement problem that creates. It does not announce a solution, because there is no honest way to claim one.

What changed

Automated crypto promotion is not new. What changed is the cost and the detectability of the output.

The methods that worked — near-duplicate detection, phrase blocklists, obvious template structure — worked because generation was expensive, so operators reused text. That constraint is gone. Content-based detection has lost most of its power, and the remaining signal has moved to behaviour and structure rather than to the text itself.

What still carries information

None of these identifies generated content on its own. Each is a distributional oddity that is consistent with coordination and also has innocent explanations. Together they are suggestive; individually they are not evidence.

• Timing structure. Organic attention arrives with human rhythm — time zones, working hours, bursts around news. Distributions that are too uniform, or too tightly clustered around a moment with no corresponding event, are anomalous. Scheduled legitimate marketing looks similar.
• Account age and history. Sudden participation from accounts created in the same window with no prior activity is a classic pattern. It is also what a genuine new community looks like.
• Engagement shape. Organic conversation is heavy-tailed: a few posts get most of the response and most get almost none. Synthetic activity often produces suspiciously even engagement, because the amplification is applied uniformly.
• Network topology. Genuine communities form clustered, irregular interaction graphs. Coordinated networks tend toward higher symmetry and reciprocity than organic ones.
• Divergence from independent measures. The most useful check is not internal to the social data at all. Social activity that rises sharply with no corresponding movement in turnover, holder counts or news flow is the pattern most worth examining. This is the reasoning behind comparing social attention against trading activity in WSD.

Organic and synthetic are not two categories

Framing the problem as "real versus fake" makes it look more tractable than it is. The realistic spectrum:

1. A person writes their own post.
2. A person uses a model to help phrase a view they hold.
3. A person schedules model-drafted posts they approved.
4. A paid promoter posts genuine-sounding content on instruction.
5. An autonomous agent posts on a topic within a mandate its operator set.
6. A coordinated network posts to manufacture the appearance of interest.

Only the last is unambiguously manipulation. The middle cases are ordinary and largely benign — using a writing tool does not make an opinion insincere. Any classifier drawing a hard line lands somewhere inside that gradient, and where it lands is a policy decision, not a technical finding.

Agents as market participants, not just noise

A second development is worth separating from manipulation entirely. Autonomous agents increasingly consume market data — monitoring news, comparing assets, summarising research — and some of them publish their conclusions.

Output from such an agent is not fraudulent. It may be more consistent than a human analyst. But it changes what a social metric measures: when agents act on published data and post about it, and those posts feed back into the data other agents read, the conversation is partly a function of the metrics rather than an independent observation of interest.

This is the reflexivity problem described in social intelligence, with a shorter feedback loop. It is also why Moonboard treats structured data access for agents as a design question rather than only a product one — see developers.

What this means for the metrics

Stated plainly, because these are the operating constraints on every social number the platform publishes:

• No authenticity filter is applied. Social volume, interactions and dominance count what was collected. Coordinated activity enters exactly like organic activity.
• Small assets are most exposed. Manufacturing a visible social footprint costs little when the baseline is a few hundred posts a day. Moving a large asset's metrics is far more expensive.
• The Moonboard Score is directly sensitive, because inflating social dominance raises the social-to-market ratio the score is built around.
• Divergence beats levels. Cross-checking a social reading against turnover, news flow and price behaviour is more informative than any single social number.

Open questions

• What fraction of measured social volume for mid-cap assets originates from coordinated activity? Unknown, and not currently measurable without ground truth.
• Do the behavioural signals above survive as operators adapt to them, or does publishing them accelerate evasion?
• Can a useful confidence interval be attached to a social metric, rather than a point estimate that implies precision it does not have?
• Is there a defensible ground truth to build against — for instance platform-disclosed automation labels — and how biased would a sample drawn from it be?