AI Slop: What Merriam-Webster’s 2025 Word of the Year Means and How to Avoid Low‑Quality AI Content

How to spot AI slop online

There’s no single reliable detector, and AI detection tools are fallible. Still, a few patterns help:

• Superficial coherence – Smooth sentences that say little, contradict themselves, or avoid specifics.
• Recycled templates – Repetitive phrasing, listicles with padded points, generic examples.
• Unverifiable claims – No citations, broken links, or links that don’t support the statement.
• Visual glitches – Inconsistent lighting, anatomy errors, or mismatched brand assets in ads.
• Shallow personalisation – Vague references to “your needs” without domain nuance.

Use these as signals, not proofs. High-quality content can be AI-assisted, and low-quality content can be human-made.

Avoiding slop in your workflow: a practical checklist

1) Start with purpose, audience and constraints

• Define the user outcome (what should this help someone do?).
• Set scope: what’s in/out; what data is authoritative; what is unknown.
• Choose the right tool: not every task needs a model.

2) Ground models in sources and cite them

• Use retrieval augmented generation (RAG) – fetch relevant documents and have the model answer from those, not from general training data.
• Require citations or links to primary sources. Prefer official docs, model cards and research papers.
• Block the model from guessing when sources are missing. “Not disclosed” is better than a confident error.

3) Keep humans in the loop

• Editorial review for accuracy, clarity and tone, especially on regulated topics.
• Domain experts sign off on technical or medical/financial content.
• Track edits to see where the model consistently falls short.

4) Label synthetic media and adopt provenance

• Be transparent when AI assisted. Clear labelling builds trust.
• Adopt content credentials and provenance standards where feasible (e.g., C2PA) so assets carry creation history.

5) Measure quality, not just throughput

• User outcomes: task success, time saved, complaint rate.
• Engagement quality: dwell time, return visitors, context-appropriate conversions.
• Quality gates: block publication if sources are missing or tests fail.

6) Don’t outrun UK compliance

• Advertising: the ASA/CAP Code prohibits misleading claims and unclear advertorials, including AI-manipulated imagery.
• Privacy: UK GDPR applies if you process personal data; do a DPIA for new AI deployments.
• Online harms: the Online Safety Act empowers Ofcom to set duties of care for platforms; expect stronger pressure on misinformation and transparency.

7) Use AI to raise the floor, not lower the bar

• Draft, then refine – use AI for outlines, data extraction and factual scaffolding, not finished prose.
• Automate the boring bits: deduping, summarising long transcripts, generating test cases.
• Reserve human time for synthesis, judgment and voice.

For UK teams: governance and procurement basics

• Document your model choices, prompts and guardrails. Treat prompts as configuration, not magic.
• Prefer vendors with clear data handling, evaluation reports and opt-out from training on your inputs.
• Create an escalation path for risky outputs (defamation, bias, medical/financial claims).
• Run regular red-team tests for hallucinations, bias and prompt injection. Patch patterns, not one-offs.

Not all AI is slop: where it genuinely helps

Used well, AI accelerates useful work without sacrificing accuracy:

• Summarising long documents while linking to sections for verification.
• Structuring unstructured data (e.g., extracting fields from PDFs) with confidence thresholds.
• Drafting code comments, unit tests and migration plans subject to code review.
• Generating variations for A/B tests, with human curation.

If you’re operationalising content generation at scale, build quality checks into the pipeline, not as an afterthought. For example, when connecting models to spreadsheets or workflow tools, pair automation with validation and clear ownership. See my walkthrough on how to connect ChatGPT and Google Sheets for practical integration tips you can adapt responsibly.

Culture matters: setting a higher bar

“Like slime, sludge and muck, slop has the wet sound of something you don’t want to touch.”

That line from the announcement sums it up. If teams reward speed and volume alone, you’ll get slop. If you reward accuracy, usefulness and clear sourcing, AI becomes a force multiplier rather than a noise machine.

The internet is being reshaped by the economics of generation. We can’t stop that, but we can choose to publish less – and better.

Read the discussion

• Reddit thread: ‘Slop’ is Merriam-Webster’s 2025 Word of the Year
• Coverage: CNET article