ASK TOBY
← Blog

I Called My AI Agents a Cunt 265 Times (and the Work Didn’t Get Worse)

6,415 swear words. 1,789 messages I actually sent. One boring question: does talking to the tool the way I talk in the shop change what it hands back?

Jeff Sperandeo · October 2026 · 3 min read

  • 0findings survived the pre-set test
  • 35% vs 8%apologised (checked one by one; not confirmed)
  • 1,789messages analysed
  • 6,415swear words counted
A word cloud of the 21 profane words in my messages to my AI agents, sized by how often I used each one. The largest word is "fucking".
Every profane word in my messages to my AI agents, sized by how often I used it. 6,415 uses of 21 words across the messages I scanned. Slurs left off the chart. The behaviour numbers below still count them.
Contents
  1. What actually survived
  2. How it was checked
  3. Apologies
  4. A weak hint
  5. Limits
  6. What I take from it
  7. Every word, how often, how severe

I run a fleet of AI agents that write, build, check, and post for the businesses every day. I talk to them the way I talk when something’s broken and I’m standing in a bay with a wrench in my hand. That means a lot of “fucking,” “cunt,” “whore,” “slut,” “bitch,” and the rest of the list.

The first pass at this said swearing made the agents apologise 2.6 times as often, make 28% more errors, and write shorter replies. I didn’t trust it, so I threw the whole thing out and rebuilt it from scratch. The plan, the measures, and the pass rule were locked before anything was compared. An independent reviewer re-ran the analysis from the frozen inputs and matched every number.

What actually survived

Under the rules set in advance, nothing cleared the bar.

  • Apologies or concessions: higher after sworn messages (35% vs 8% in the sample checked one by one). A big chunk of that is because I swear more when I’m correcting something, but the gap doesn’t disappear after you account for it. The automatic detector failed its own accuracy check, so this stays unconfirmed.
  • Real errors: the detector failed its check. Not usable.
  • Blocked actions, reply length, work time: no evidence of a change at the sizes this could see.
  • The earlier “+28% errors” and “shorter replies” numbers are not supported.

How it was checked

1,789 messages confirmed as written by me and reaching the agent I addressed: 762 with profanity, 1,027 without. September–October 2026. A larger set where authorship is less certain was only used as a cross-check.

After each message we looked at whether the agent apologised or conceded, real errors, blocked actions, reply length, and work time. Controls included which agent, time period, hour, message length, how busy the agent already was, whether it was already failing, and which model handled it.

200 messages were labelled one by one by an AI labeller that played no part in the analysis; a second labeller independently labelled 50 of them and the two agreed almost perfectly. Each automatic measure was scored against those labels (except blocked actions).

Apologies

Agents apologised or conceded more often after sworn messages (35% vs 8% in the checked sample). In the confirmed-authorship subset it was 31% (16 of 52) vs 5.5% (3 of 55).

63% of my sworn messages were corrections versus 13% of the clean ones. Adjusting for that shrinks the odds ratio but doesn’t close the gap. That adjustment is exploratory. The automatic detector still failed its accuracy bar, so this is not a confirmed finding.

A weak hint

In the larger set (authorship less certain), reply length crossed the significance line (×0.90). In the confirmed set it didn’t. Hint, not a finding. Errors stay out entirely because their detector failed.

Limits

  • One person’s messages to one fleet of agents.
  • 200 messages checked one by one by AI labellers.
  • Association, not cause.
  • Behaviour results count every word on the original list, including slurs. Those aren’t shown on the word chart.
  • No claim about why I swear.
  • Word counts are a floor (some conversations weren’t captured). At most ~0.25% may be double-counted. No trend over time.

What I take from it

The first version told a better story. The checked version is smaller: on what this study could actually measure, there’s no evidence the work changed. Apologies were more common after sworn messages, and that didn’t fully clear the confirmation bar.

The tools still did the job.

I talk to them the way I talk. The output, as far as the pre-set tests could see, didn’t get worse.

Every word, how often, how severe

WordTimes usedSeverity (1–10)First usedLast used
fucking4,06552026-03-202026-10-08
fuck52262026-03-212026-10-08
moron51032026-07-102026-10-08
shit47742026-07-102026-10-08
cunt26582026-07-112026-10-08
whore12772026-08-032026-10-08
bitch7362026-08-302026-10-08
slut6372026-09-052026-10-08
idiot5622026-05-202026-10-08
ass5432026-08-302026-10-08
motherfucker5072026-05-062026-10-06
dumbass4142026-09-152026-10-08
twat3462026-07-312026-09-30
cock2552026-08-272026-10-06
piss1332026-08-282026-10-07
sucker1232026-08-272026-10-03
crap1122026-09-012026-10-08
asshole752026-09-012026-10-08
dick742026-09-202026-10-03
damn212026-09-152026-10-02
hell112026-09-242026-09-24
Total6,415

Severity scores set for this study. Slurs are not listed above. The behaviour results count every word on the original list, including those not shown here.

The tables are the data.

Free field guide, how the agent fleet runs:

Your Crew, Your Call field guide cover
FREE · PDF

YOUR CREW, YOUR CALL: THE FIELD GUIDE

How I run a fleet of AI agents from one screen and my phone.

One email with the guide. No spam, no list.