The Quiet Revolution

A draft can sound right before it is right. That is the useful tension in OpenAI’s release of GPT-4.5 this week: a model that may be easier to write with still leaves us responsible for what the finished communication says.

OpenAI introduced GPT-4.5 on February 27 as a research preview. The company reports improvements in natural conversation, following user intent, and writing. Those are promising areas for workplace use, but launch tests are an invitation to evaluate a tool in our own work.

What Launched, and What Did Not

GPT-4.5 belongs to OpenAI’s effort to scale pretraining and post-training. The announcement explicitly distinguishes it from reasoning models such as o1 and o3-mini, which spend additional computation working through a response. Calling GPT-4.5 a new reasoning model obscures that distinction.

Availability also matters. The February 27 release starts with ChatGPT Pro and paid API access, with other ChatGPT tiers scheduled to follow. As of March 2, a planned rollout is not the same as availability to every employee in an enterprise.

The company’s description of improved “EQ” refers to conversational behavior and user preferences. It is not evidence that the model feels empathy, understands a colleague’s experience, or learns permanently from every conversation. A useful tone and a trustworthy statement still need to be assessed separately.

What the Writing Study Measures

There is evidence that AI-assisted writing is spreading. A February 13 preprint by Weixin Liang and colleagues examines consumer complaints, corporate press releases, job postings, and United Nations press releases through September 2024.

For corporate press releases, the authors estimate that up to 24% of text is attributable to language-model assistance. That is narrower than saying nearly half of all corporate communications are AI-written. It concerns a particular corpus and a statistical estimate of assistance, rather than a count of whole documents written entirely by machines.

Population-level analysis also cannot tell us that a particular employee used AI. And this study does not establish that the people using it gained time for more creative work, improved their decisions, or became more productive. Those questions require different evidence.

What it does give us is a reason to examine the writing process. If these tools are becoming part of ordinary communication, responsibility for the words needs to remain visible.

Test the Whole Communication Task

Consider an internal project update. The task is not simply to produce three polished paragraphs. It is to tell colleagues what changed, what remains uncertain, and what they need to do.

An AI-assisted draft could help organize the source notes. It could also turn an unresolved issue into a reassuring sentence, omit a qualification, or make a proposed date sound agreed. Each mistake creates work for someone farther down the chain.

For a small comparison, give a reviewer the same approved facts and ask for two versions: one produced through the existing process and one with AI drafting assistance. Have the reviewer check whether each version preserves the facts, makes the requested action clear, and sounds appropriate for its audience.

Record the time spent checking and correcting, as well as the time spent drafting. If the tool saves ten minutes of composition but introduces fifteen minutes of verification, the relevant task did not get faster. If it helps someone express a difficult point clearly without changing its meaning, that is a benefit worth identifying.

Keep the Person Behind the Words

Authenticity is partly a question of style, but it is also a question of ownership. A message should reflect what its sender knows, intends, and is prepared to do. A warmer sentence is a problem if it promises support that nobody has agreed to provide.

Use the draft to improve expression, then return to the source facts. Confirm the commitments. Restore uncertainty where the evidence is uncertain. Read the result as the person who will receive it.

GPT-4.5 gives us another opportunity to test how AI can help with that work. Its launch does not settle how much work will change or what the next model will deliver. The immediate question is smaller and answerable: did this tool help us send a clearer, more accurate message that a person can stand behind?