Is it wrong to use AI to write posts at all?

No. It is wrong to publish something nobody thought about. AI is excellent at research, interrogation, structure, and mechanical checks; it is unreliable as the author of the load-bearing claim, because it has no stake in the claim being true.

Can readers really tell?

Increasingly, yes, and where they cannot, they can tell that nothing was learned. LinkedIn shipping a report button and a comment-spam enforcement program is evidence that reader detection is now a product-level concern.

What about detectors?

Statistical AI detectors are unreliable in both directions and should not be used to accuse anyone. Check for the absence of first-party evidence instead. That is measurable and it is what actually matters.

How do I apply this to a team of ten writers?

Make the brief the artefact you review, not the draft. If briefs are approved with real evidence in them, draft quality mostly takes care of itself, and review time drops by more than half.

Journal

Content operations

The Anti-Slop Standard

Slop is not a style problem. It is a process problem with a style symptom, and you can fix it with rules a machine can check.

Alex Cinovoj TechTide AIFebruary 10, 20263 min read

Key takeaways

Ban the tells: em-dash cadence abuse, "in today’s landscape", triads, and hedged non-claims.
Require first-party evidence (a number, a name, a date, or a failure) in every piece.
Enforce mechanically. A rule nobody can check is a preference, not a standard.
Put the gate before generation so the model cannot produce a polished nothing.

The word slop won because it is accurate. Slop is not badly written; that is what makes it insidious. It is competently written text with no load-bearing content, fluent, structured, agreeable, and completely empty. LinkedIn now ships a button for reporting it, which means the market has priced it in.

Most teams respond by editing harder. That fails, because editing operates on sentences and slop is a defect in the brief.

The tells, and why they show up

Generated prose has a fingerprint because the objective that produced it is fluency, not commitment. Fluency under uncertainty produces a specific set of moves.

The triad. Everything comes in threes: "clear, concise, and compelling." It reads as rhythm and functions as padding.
The hedged claim. "This can often help teams to potentially improve." Four qualifiers and no assertion.
The false stakes opener. "In today’s rapidly evolving landscape…" A sentence that could precede any article ever written.
Symmetry. Every section the same length, every list the same shape. Human argument is lumpy; the important part gets more room.
Topic-sentence stacking. Each paragraph restates its own heading. Delete the first sentence of every paragraph and see how much survives.

What a real standard requires

Bans alone produce careful emptiness. The standard has to require things too, and the requirements should be checkable.

One proprietary fact. A metric from your own system, a customer situation with details, a decision and its date. If the brief cannot supply one, the piece is not ready.
One position. A sentence a competent peer could disagree with in public.
One cost. What the recommendation breaks, who it annoys, when it is wrong.
Attribution for every borrowed claim. Named, linked, and quoted briefly, not laundered into your own prose.
A named author. A person accountable for the claims, not a brand voice.

Enforce it before generation

This is the part most tools get backwards. If the gate runs on the output, the model will happily produce a beautifully structured piece with nothing in it, and a tired human at 6pm will approve it. The gate belongs upstream: the brief must contain the proprietary fact, the position, and the cost, or nothing generates at all.

That is exactly what our pre-flight gate does, four checks, all green or the run does not start. It is a deliberately annoying design. The annoyance is the feature: it converts "we should write something this week" into "we know something worth saying."

The rules a machine can actually check

Some of the standard is judgement. A surprising amount is not. Our checker rejects a draft when:

- banned_phrases matched (in today's, landscape, delve, tapestry, game-changer)
- claim_density < 1 verifiable claim per 150 words
- hedge_ratio > 0.08 (qualifiers / total words)
- zero first-party entities (no metrics, dates, or named systems)
- paragraph length variance < threshold (suspicious symmetry)
- no outbound citation for any borrowed statistic

None of these measure quality. All of them measure the absence of the specific failure mode. That is the right ambition for automation: catch the floor, leave the ceiling to the human.

Automate the floor. Never automate the ceiling.

The honest trade-off

A standard like this cuts output. Teams that were shipping five posts a week land at two, and the first month feels like regression. What comes back is that the two get read, quoted, and replied to by people who matter, which was the actual goal that volume was proxying for.

If your reporting rewards published count, this will not survive contact with your dashboard. Change the metric first. See attribution for content for what to measure instead.

The scoring rubric behind this standard lives in the open on alexcinovoj.com (opens in a new tab), and we run it against real client work at TechTide AI (opens in a new tab). Copy it, argue with it, tighten it for your own category.

“You cannot edit your way out of a draft that had nothing to say. The gate has to be before generation, not after.”

Alex Cinovoj, TechTide AI

Questions people actually ask

Is it wrong to use AI to write posts at all?: No. It is wrong to publish something nobody thought about. AI is excellent at research, interrogation, structure, and mechanical checks; it is unreliable as the author of the load-bearing claim, because it has no stake in the claim being true.
Can readers really tell?: Increasingly, yes, and where they cannot, they can tell that nothing was learned. LinkedIn shipping a report button and a comment-spam enforcement program is evidence that reader detection is now a product-level concern.
What about detectors?: Statistical AI detectors are unreliable in both directions and should not be used to accuse anyone. Check for the absence of first-party evidence instead. That is measurable and it is what actually matters.
How do I apply this to a team of ten writers?: Make the brief the artefact you review, not the draft. If briefs are approved with real evidence in them, draft quality mostly takes care of itself, and review time drops by more than half.

Sources & further reading

Written by

Alex Cinovoj, Founder, TechTide AI

Alex Cinovoj builds ClawLI at TechTide AI, a fleet of specialist agents that research, draft, schedule, and grade LinkedIn content with a human approving every send. He writes about what breaks when you put agents in front of a real audience.

alexcinovoj.com techtideai.io See the pre-flight gate