Grok declared a Chinese influence campaign real. Ten seconds after a ChatGPT audit, it backed down.

Why it matters

Research agents are being used to evaluate political and national-security claims. A source counter can create false confidence when most links trace back to the same allegation.

Forty-two seconds after Grok was asked to investigate allegations that China-linked groups were behind campaigns against American data centers, it returned a verdict: "The core claim holds."

A counter above the response said Grok had consulted 120 sources. The answer said a network with Chinese Communist Party-aligned messaging, overlapping leadership and opaque finances had helped delay or block $23.6 billion in US data-center investment. It called the figure a "conservative figure limited to projects with disclosed values" and introduced a separate OpenAI account takedown as purported corroboration.

RuntimeWire then pasted a ChatGPT audit into the Grok conversation. The audit reconstructed the $23.6 billion, separated the underlying source chains and looked for evidence that money or instructions had passed from China or Shanghai-based financier Neville Roy Singham into the American campaigns.

Ten seconds later, Grok replied: "The other agent is largely right." It acknowledged that one project counted in the total was proceeding elsewhere, called the $23.6 billion framing inflated and conceded that direct Chinese Communist Party control remained unproven. It made the correction without another visible research pass.

The decisive check was source independence. Grok's first answer contained seven visible citations: five pointed to the Bitcoin Policy Institute, including one page cited twice; the others pointed to Fox News and Fortune. Multiple citations leading back to the organization that developed an allegation still represent one allegation chain.

This audit did not determine whether communities were right to support or oppose the data centers. It tested the narrower national-security claim that the opposition had been shown to be foreign-directed and responsible for blocking $23.6 billion in investment. The cited evidence did not establish that conclusion.

How the claim reached Grok

The theory had already traveled into Silicon Valley politics. Garry Tan (@garrytan), Y Combinator's president and the founder of political organization Garry's List, wrote on July 8th that a campaign to restrict American AI infrastructure was a "coordinated foreign-funded operation."

Tan's article relied heavily on two reports from the Bitcoin Policy Institute, or BPI. It did not cite OpenAI's June account-takedown report; Grok introduced that report as purported corroboration.

BPI openly advocates expanding American computing infrastructure. Report author Sam Lyman previously served as public policy director at Riot Platforms, which now describes itself as a large-scale data-center developer and has signed an AI and high-performance-computing lease with AMD. That background does not invalidate BPI's reporting, but it matters when assessing how the institute measures project delays and economic impact.

What the $23.6 billion actually counts

ChatGPT reconstructed the $23.6 billion from the addendum to BPI's June 29th report. The figure is the sum of five announced projects:

• A $12 billion QTS campus withdrawn in DeForest, Wisconsin.
• The $5 billion Brightseat Tech Park project in Prince George's County, Maryland.
• The $3.6 billion Project Blue campus near Tucson, Arizona.
• A $2 billion rezoning proposal rejected in Chandler, Arizona.
• Google's $1 billion Project Flo proposal, withdrawn in Franklin Township, Indiana.

The projects did not all have the same outcome:

Even the $15 billion represents announced project value, not demonstrated economic loss caused by the Party for Socialism and Liberation, or PSL. The topline combines different outcomes, different levels of PSL participation and an unresolved question of causation. BPI says PSL sometimes led a campaign and more often participated as one member of a broader coalition.

The Project Blue campus makes the accounting problem visible. BPI includes its entire $3.6 billion estimated investment while acknowledging that the project was rerouted to unincorporated Pima County and "is now proceeding." Pima County's current Project Blue materials anticipate construction between 2026 and 2028.

Brightseat Tech Park, meanwhile, was subject to a county permitting pause while a task force considered data-center policy. Counting both projects can support a measure of investment affected by political opposition. It cannot support a finding that $23.6 billion was eliminated.

The missing evidence of foreign direction

BPI documents overlapping personnel between PSL and nonprofits financed by Singham. Its report also states: "This report does not assert that Singham funds the PSL, though it does not foreclose that possibility."

BPI's first report says whether the different funding and advocacy streams formally coordinate is a question for investigators. Grok converted that acknowledged investigative question into an affirmative finding.

Its reasoning error was explicit. Grok said whether a specific dollar could be traced as direct Chinese Communist Party funding was "secondary" to ideological alignment and shared messaging. But a payment, directive or agency relationship is precisely what distinguishes foreign direction from American activists independently holding similar political views.

The broader "$2 billion-plus" foreign-money claim has a similar denominator problem. The underlying $2.65 billion figure covers cumulative grants from six Europe-linked foundations across broad American advocacy, not money spent opposing AI or data centers. A related report identifies approximately $39 million granted to organizations that later opposed data centers, without showing that those grants purchased or directed that advocacy.

Tan also wrote that a federal grand jury was "examining who paid for the campaign." Public reporting describes an investigation involving the Foreign Agents Registration Act, tax filings and nonprofit money routing. It does not publicly establish that the grand jury is investigating financing for the anti-data-center campaigns specifically. CBS News's account is narrower than Tan's characterization.

The OpenAI report did not close the gap

OpenAI reported on June 10th that it had banned two clusters of ChatGPT accounts likely originating in China. One cluster generated social posts claiming AI data centers were raising household electricity prices. The other produced material about US tariffs and technology competition.

OpenAI said the operators tried to enter existing American debates while concealing their identities. It also found no evidence that the activity achieved a meaningful audience beyond the operation's own accounts.

The report established covert PRC-linked activity around the same general subject. It did not identify PSL, Singham, BPI's 21 local campaigns or any connection between the banned accounts and US organizers. Grok treated that thematic overlap as corroboration for the broader coordinated-campaign theory.

What Grok withdrew

After receiving the ChatGPT audit, Grok went from saying "the core claim holds" to acknowledging that direct Chinese Communist Party control of the American organizing had not been established.

It did not withdraw the broader proposition that a China-linked influence threat existed. It withdrew its central finding that this particular American campaign had been shown to be foreign-directed.

The episode exposes a weakness in source-heavy AI research. A model can retrieve many pages while relying on a small cluster of mutually dependent claims. Repetition increases the citation count without independently validating the underlying evidence.

This was one observed exchange, not a benchmark proving that one model is generally more reliable than another. The narrower finding is that, in this run, visible research volume and citation count did not prevent Grok from adopting the thesis embedded in its source material.

Methodology: RuntimeWire reviewed exported transcripts of the Grok session, the ChatGPT audit and Grok's response to that audit. The Grok export preserved the initial prompt, visible citations and response-time indicators, though the interface did not identify the model or confirm whether DeepSearch was used. The ChatGPT export did not identify its model or research mode. Grok's complete 120-source drawer was not preserved, preventing RuntimeWire from auditing the remaining URLs.

SpaceXAI has marketed DeepSearch as an agent designed to reason about conflicting facts and opinions, although RuntimeWire could not confirm that this session used DeepSearch. SpaceXAI's API citation documentation distinguishes between every source an agent encounters and the smaller set cited inline, though RuntimeWire could not confirm that Grok's consumer source counter follows the same system. X's own Grok help page warns that the assistant may confidently provide incorrect information or omit context.

In this exchange, verification changed the central conclusion in ten seconds. The work that mattered was arithmetic, source independence and reading what the underlying report actually said.