We’ve all sat through them: speeches that miss the mark so widely you wonder how the speaker was selected and invited in the first place. I’m not talking about your CEO or the chair of your board. They get selected by virtue of position and, if they’re good at it, you don’t mind listening to them talk. If they suck at it, you listen anyway because it’s your job. But when you go to a conference, or a board of trade event, or a chamber of commerce get together, and the featured speaker is simply terrible, you probably feel cheated out of several hours you could have spent in enjoyable alternative realities.

The classic mistakes are easy to dredge up from our painful memories: reading from a script with zero eye contact with the audience; talking above or below the level of knowledge of the audience, or their appetite for technical detail; sexist, racist or cultural jokes; umming or losing their place or train of thought; lacking structure that helps the audience follow along; or victim body language signaling a distinct discomfort with the whole endeavour.

It reminds me of the old movie, Planes, Trains and Automobiles where Steve Martin gives it to John Candy in an explosion of listener frustration.

Recently, bad speeches have taken on a new flavour. All the old performance issues are still there, the stumbling, lack of eye contact and so on, but the biggest change is in the content itself. What I’m noticing is content that is cliché ridden, patronizing, and overly pedantic. It’s as if ChatGPT and Dale Carnegie had a baby.

People say they can pick out something written by ChatGPT and Prasanth Mannava posted this handy infographic to help you get better writing out of your favourite AI by telling it what to do and not do. It also serves to summarize the AI-written tells.

There are more tools coming online which purport to identify AI generated content, generally appealing to places where plagiarism is an issue. Here’s a nifty excerpt from Scispace on language tells to watch or listen for.

Look for AI Writing Red Flags in Language

AI often produces text with specific linguistic patterns. Humans show idiosyncrasies, while AI tends toward cautious, polished uniformity—especially at low “temperature.” Use these cues as triage prompts, then verify with other methods.

* Repetitive structure: similar sentence lengths; recurring clause templates (“While X, Y also Z”).

* Vague generalities: filler phrases (“in today’s world,” “it is important to note”) standing in for concrete detail.

* Unnatural transitions: abrupt topic shifts; connective tissue without substance.

* Hallucinations: confident but nonexistent references, wrong dates, or mismatched stats.

* Flat voice: an oddly consistent tone that resists personal perspective, hedges excessively, or avoids specifics.

Example 1: A post repeats “it’s important to note” five times—triage as AI‑like.Example 2: An “experience” essay lacks concrete places, dates, or sensory detail—ask for corroboration.

Analyze Tone and Style Consistency

Human writing typically maintains a coherent voice shaped by experience and audience; AI may drift as prompts shift. Stylometry and simple side‑by‑side checks reveal inconsistencies.

* Voice drift: sudden switches between formal and casual, or between British and American spelling conventions.

* Passive overuse: “was conducted,” “were analyzed” hiding agency across paragraphs.

* Emotional flatness: descriptions that lack sensory detail where it would be natural.

* Compare to known samples: authors’ prior work, emails, or drafts provide a baseline.

* Stylometry tools: optional analysis of function‑word usage and sentence rhythm.

Example 1: An essay opens with lyrical scene setting, then shifts to generic corporate tone—ask for earlier drafts.Example 2: A research summary alternates “color”/“colour” and date formats, suggesting multi‑source or generated edits.

When it comes to delivering this kind of writing out loud, the soullessness becomes even more obvious, as does the complete departure from how a person you know normally speaks. In the case of a complete stranger, you might notice it’s a complete departure from how anyone speaks!

Personally, I rely on an amorphous kind of knowing, reminiscent of Jacobellis v. Ohio, an obscenity court case dating back to 1964 where one of the Supreme Court justices, Justice Potter Stewart opined that only “hard core pornography” constituted obscenity. He went on to say, in a much quoted summary, that he could not define hard core pornography, nor obscenity but, “I know it when I see it”. And there it is.

The fear is AI generated content is becoming so pervasive I may well accuse someone of using it to write their speech, only to discover they did write it themselves, based on immersion in what now passes for good writing. Life imitating tech, God help us.

So many people already struggle to speak in front of an audience and nerves are tough to disguise, although they can be trained. The short-cut of having AI write your speech, then expecting your performance to somehow rise above vacuous content, is a mug’s game. We would rather hear from you, in your own words, no matter if they’re not perfect, no matter if you are nervous. Respect the audience, assume positive intent, be human.

Therefore, consequently and in conclusion…ha! Gotcha!

AI is already putting ideas into our heads, now words into our mouths. Once we begin to move our bodies according to its instructions, we will have become meat puppets, animated by LLMs using social media and legacy media platforms. I’d rather take my directions from the Bangles, circa 1986.

Until next time, get up on that stage and show those machines how it’s done!

Talking With Friends, Sharing the Load is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit joannapiros.substack.com/subscribe