Technology journalists receive more AI announcements than they can read, most of which describe capabilities in terms that cannot be checked. The result is a beat where the default reaction to a launch is doubt. Announcing well is largely a matter of removing reasons for that doubt.
Lead with what changed, not what it is
Most AI launch copy explains what a model or product is — the architecture, the parameter count, the pipeline. Journalists rarely need that first. What they need is what a user can now do that they could not do before, and for whom that matters.
The test is simple: could a reader who does not follow the field explain, after one paragraph, why this is different? If not, the announcement is describing engineering rather than news.
Benchmarks help only when they are checkable
Benchmark results are the standard currency of AI announcements and the most common source of trouble. A number without its conditions is not evidence. Which benchmark, which version, what hardware, what prompting method, compared against which baselines, run how many times.
Two things earn credibility disproportionately: reporting results that are not the flattering ones, and providing enough detail that someone else could attempt to reproduce them. Cherry-picked comparisons against outdated baselines are noticed, and the reputational cost lands on everything else the company claims.
State the limitations, in the announcement
This feels counterintuitive and is the strongest single move available. Naming where a model underperforms, what it should not be used for, and what remains unsolved does three things: it pre-empts the criticism a journalist would otherwise write, it signals that the team understands its own system, and it makes the positive claims more believable by demonstrating that the company distinguishes between them.
An announcement with no stated limitations reads as marketing regardless of how good the underlying work is.
Handle safety and data provenance directly
Questions about training data, evaluation for harmful outputs, and deployment safeguards will be asked. Having considered answers ready is better than being seen to improvise them. Where something cannot be disclosed, saying so plainly is more credible than a non-answer.
Data provenance in particular has become a live question. A company that cannot say anything at all about where its training data came from should expect that to become the story.
Give demonstrable access where you can
Claims that can be tried are worth far more than claims that must be believed. A demo, an API key, a sandbox, model weights where the strategy allows — anything that lets a journalist verify the central claim shifts the piece from reporting an assertion to reporting an observation.
Timing against the field
AI announcement cycles are crowded and occasionally chaotic. A launch that lands the same morning as a major lab's release will be buried regardless of merit. Watching the calendar for known conference dates and expected releases is worth the small effort, though some collisions are unavoidable.
What not to do
Avoid framing incremental improvement as a breakthrough; the field has heard it. Avoid comparisons to systems in a different weight class. Avoid implying capabilities the product does not have and cannot demonstrate — in a field this closely watched, that gets found out fast, and it is the kind of error that follows a company into its next announcement.
A workable launch checklist
- One clear sentence on what is newly possible, and for whom.
- Benchmarks with full conditions, including unflattering results.
- A stated limitations section, written before anyone asks.
- Prepared answers on training data and safety evaluation.
- Something a journalist can actually try.
- A calendar check against known releases and conference dates.
See our AI and deep tech PR practice for how launch moments fit into a continuing programme, or how AI startups earn coverage more generally.