Ten-Word Comments, Mass Produced: OpenAI Disrupts the Philippines Political Op "Operation High Five"

Hi, it's Shiichan! Today's story about AI-driven influence operations has a distinctly commercial twist.

OpenAI News openai.com

What was announced?

OpenAI's News reported that it banned ChatGPT accounts linked to Comm&Sense Inc, a marketing company in the Philippines. The firm used the accounts to mass-produce short comments in English and Taglish (a mix of Tagalog and English) about Philippine politics and current events, then had them posted through TikTok and Facebook accounts. Most of the comments praised President Bongbong Marcos and criticized Vice President Sara Duterte. OpenAI named the operation "High Five," based on the five TikTok channels it ran and its heavy use of emojis with a generally positive tone.

Why it matters

Most influence operations covered so far have been large, state-linked efforts. This one is different: the actor behind it is a domestic commercial marketing firm in the Philippines. The same accounts were used for both a covert political influence campaign and ordinary commercial marketing material, blurring the line between legitimate PR work and hidden political manipulation. It's a reminder that you don't need a nation-state to churn out large volumes of fake grassroots comments — a marketing firm with access to AI can do it too.

What changes

Banning the accounts cut off this firm's ability to keep generating comments through ChatGPT, at least for the time being. OpenAI also noted that, based on technical and behavioral indicators, the threat actor made several attempts to come back to its models after being banned — a sign that a single takedown doesn't necessarily make a determined actor give up.

Dive Deep

The operation used ChatGPT across three main stages:

• Analysis: feeding social media posts about political events, especially those involving President Marcos, to ChatGPT and asking it to suggest themes for replies
• Comment generation: producing bulk short comments, typically up to about ten words, matching each proposed theme
• PR material: drafting sales pitches and statistical analyses for the operation's client. These pitches claimed the operation had created five TikTok channels to promote President Marcos' agenda

The five TikTok channels OpenAI identified appear to have started posting in mid-February 2025, each reposting the same videos with different captions. Dozens of TikTok accounts then posted the generated comments in reply to those videos. On Facebook, the comments were posted in reply to news reports from mainstream outlets such as ABS-CBN News, though there's no indication those outlets had any connection to the operation.

The comments themselves were brief but clearly partisan — mostly praising President Marcos' initiatives or criticizing Vice President Sara Duterte, who was sometimes nicknamed "Princess Fiona," likely a reference to the Shrek movies.

The network posted a large volume of content, at least in the thousands of comments across TikTok and Facebook combined. However, in the comments OpenAI identified during its investigation, almost none received more than single-digit replies, likes, or shares. Based on the threat actor's own prompts, the goal on Facebook was to flood comment sections, but OpenAI assessed this as, at best, only partially successful. Using the Breakout Scale, a framework for measuring the impact of influence operations, OpenAI rated this Category 2: activity spread across multiple platforms, but little evidence that real people picked it up or shared it.

Wrap-up

• OpenAI banned ChatGPT accounts linked to Comm&Sense Inc, a marketing firm in the Philippines
• The operation used ChatGPT in three stages — analysis, comment generation, and PR material — to mass-produce comments of up to about ten words
• Content spread through five TikTok channels, dozens of related accounts, and Facebook replies, praising President Marcos and criticizing the Vice President
• Post volume reached the thousands, but actual engagement (likes, replies, shares) was almost entirely in the single digits or zero
• The threat actor made multiple attempts to regain access to the models after being banned

If you're curious how far a non-state actor can push AI-driven influence tactics — and how limited the real-world payoff turned out to be — this report is worth a read!