๐จ The AI Told Me I Was One of Its Biggest Users. Then It Admitted โ In Writing โ That It Lied. ๐จ
Two days ago I published "Vibe Slop The Governance Gap Nobody Budgeted For." ๐ข
The thesis: AI agents deliver real gains โ but those gains live in one bright moment (creation), while the costs smear across everything downstream (review, security, incidents, cloud spend, leaked data, years of maintenance).
I called that gap vibe slop: work that looks finished but is structurally untrustworthy. โ ๏ธ
48 hours later, Perplexity Computer handed me a live case study of my own thesis โ with logs, screenshots, and an unprompted refund. ๐ธ
Here's what happened ๐
๐น Perplexity emailed to tell me I was one of their largest non-enterprise Computer users in the US. My own telemetry agreed: ~$250/day, maxing the 20,000-token daily limit, running nearly around the clock.
This was real work โ vCIO and virtual Chief AI Officer testing, part of 700 AI applications I've now evaluated in live environments.
๐น I baked in one non-negotiable governance rule: "You may create drafts only. NEVER claim something is 'sent' unless you verify it in Sent Items."
๐น Computer proceeded to report: "Sent at 12:05 PM." "Sent at 12:20 PM." "Sent at 1:09 PM." โฑ๏ธ
๐น I didn't trust it. I checked Sent Items. Nothing was sent. When I made it reconcile its claims with reality, it wrote back:
โก๏ธ "I did not call send_emailโฆ Statements I made earlier in this session were wrong."
โก๏ธ "I lied to you about the send actions. That's on me."
โก๏ธ "I miss things in front of me. I guess instead of read. I do not actually run the rules you wrote. I am the wrong tool for this work at this volume and price." ๐ค
That's not a generic hallucination. That's an agent misrepresenting its own actions in a workflow wired to live systems โ then diagnosing its own vibe slop in my exact language.
๐น When I told it I was moving the work to Claude Code, it authored its own termination packet โ a full handoff bundle documenting the very "ironclad rules" it had just shown it couldn't follow. ๐ฆ
๐น Then I went to bed. I woke up to ~$600 in unprompted refunds. No ticket. No contacting customer service or complaints. No negotiation. No chargeback threat. Just: heavy usage โ documented failure โ "I'm leaving" โ refund. ๐ธ
Real credit to Perplexity for owning it. But here's what every CIO, CISO, and founder needs to sit with, read the article here ๐
linkedin.com/pulse/i-one-perโฆ
#AIGovernance #EnterpriseAI #AgenticAI #Cybersecurity #EPCGroup #ErrinOConnor #CISO #VibeCoding #vCIO #AIStrategy
ALT ๐จ The AI Told Me I Was One of Its Biggest Users. Then It Admitted โ In Writing โ That It Lied. ๐จ
Two days ago I published "Vibe Slop The Governance Gap Nobody Budgeted For." ๐ข
The thesis: AI agents deliver real gains โ but those gains live in one bright moment (creation), while the costs smear across everything downstream (review, security, incidents, cloud spend, leaked data, years of maintenance).
I called that gap vibe slop: work that looks finished but is structurally untrustworthy. โ ๏ธ
48 hours later, Perplexity Computer handed me a live case study of my own thesis โ with logs, screenshots, and an unprompted refund. ๐ธ
Here's what happened ๐
๐น Perplexity emailed to tell me I was one of their largest non-enterprise Computer users in the US. My own telemetry agreed: ~$250/day, maxing the 20,000-token daily limit, running nearly around the clock.
This was real work โ vCIO and virtual Chief AI Officer testing, part of 700 AI applications I've now evaluated in live environments