Microsoft Word has spent the past few years getting increasingly acquainted with AI. Copilot can already draft text, rewrite passages, summarize documents, and answer questions about whatever you have open. Microsoft has even been rolling out more advanced editing capabilities that let Copilot make broader changes directly inside a document.

Now there is another AI sitting in Word. OpenAI launched ChatGPT for Word, via a plug-in, on September 17, putting a ChatGPT sidebar directly inside Microsoft's word processor. It works across ChatGPT's plans, including Free, although your normal ChatGPT usage limits still apply.

That creates a slightly peculiar situation. Microsoft has spent considerable effort building Copilot into Word, and now I can install its most famous AI rival in the same application. Naturally, I wanted to see which one I would actually reach for. I came away liking both, although for rather different reasons.

Installing ChatGPT in Word is straightforward, although there is one extra step compared with Copilot if Microsoft's assistant is already part of your Microsoft 365 setup. OpenAI's ChatGPT add-in is available through Microsoft Marketplace. Once installed, you can open it from the Word ribbon, sign-in to your ChatGPT account, and have the familiar chatbot appear in a sidebar.

The integration works with ChatGPT's Free, Go, Plus, and Pro plans, with the usual usage limits applying to each. The big difference with regular ChatGPT is that all conversations inside Word are separate from the ones in the regular ChatGPT app, and ChatGPT's memory does not follow you into Word either. That is worth remembering if you are accustomed to the regular chatbot already knowing your preferences.

Test 1: Planning a trip

I started with a blank document and asked each AI to respond to a request for a reusable vacation-planning template that I asked for using natural language in the sidebar.

I wanted somewhere to keep travel details, hotel information, reservations, restaurants, packing notes, and a loose daily itinerary without creating the sort of schedule that makes a vacation feel suspiciously like work.

ChatGPT produced a thoughtful setup, including the expected travel and accommodation details, but with added sections for reservation deadlines, estimated travel times, priority activities, backup plans, and ideas that had not yet made the final itinerary. Instead of filling every day with morning, afternoon, and evening slots, it left more space for a main activity and optional plans, which made the template feel less rigid.

In contrast, Copilot gave me the more conventional Word template. It neatly divided the document into sections for transportation, accommodation, daily plans, activities, dining, and packing, with plenty of space to fill everything in later. The structure was clean and immediately usable, and because Copilot is so closely integrated with Word, tweaking individual sections and reorganizing the document felt particularly easy.

Copilot was stronger at quickly producing a polished template that looked and behaved like something created specifically for Word. ChatGPT was better at considering the practical problems that tend to emerge while planning an actual vacation and building the document around them.

I ultimately preferred ChatGPT's version because it felt more flexible, although Copilot's was arguably easier to start using immediately. ChatGPT did a little more thinking about what information I might eventually wish I had included.

Test 2: Editing down nonsense

I then decided to see how they compared at editing. I pasted in an old, absurdly long set of instructions for an AirBnB I had saved years ago out of amusement, complete with information about plants, deliveries, Wi-Fi and several unnecessary paragraphs devoted to where various things were stored.

I selected a particularly rambling section and asked ChatGPT to shorten it while preserving anything somebody would actually need.

“The main problem is repetition rather than the amount of information," The AI wrote. It mentions the delivery instructions twice and explains the location of the spare supplies in more detail than someone staying for a weekend probably needs. I’d keep the Wi-Fi information, the plant-watering schedule and anything involving keys or access exactly clear. The rest can be condensed into a short ‘Around the house’ section so the useful details are easier to find.”

Copilot is hardly incapable of that conversation. It can answer questions about a document, provide summaries and respond to prompts about unclear passages, and provide references showing where information came from. But I would tend to pull up Copilot only when I already knew what I wanted it to do.

Selecting some text and telling Copilot to rewrite it produces versions that can be replaced or inserted directly into the document, and Microsoft now allows editing inside its suggestion box before accepting the result.

Which did I prefer?

ChatGPT's biggest practical advantage may simply be accessibility. OpenAI says ChatGPT for Word is available on all ChatGPT plans, including Free. Copilot availability varies according to Microsoft 365 subscription, Copilot license and organizational settings, while some of Microsoft's newer 'Edit with Copilot' features are still rolling out to eligible users.

There are good reasons to prefer Copilot. Its integration with Word is deeper, and Microsoft has built features specifically around manipulating Word documents rather than placing a general AI assistant alongside them. Depending on your setup, Copilot can also draw on Microsoft files, emails and meetings, which could matter far more than conversational style for people already living inside Microsoft 365.

For quick mechanical changes, particularly when I knew exactly what needed rewriting, Copilot's tighter relationship with Word made sense. It felt like an extension of the application rather than another destination. ChatGPT was the one I preferred when the problem was fuzzier. The two assistants overlap considerably, but they did not feel identical when I actually used them. Copilot often felt like an AI feature of Word. ChatGPT is a more fully-featured chatbot, while Copilot simply augments Word with AI features. Either is fine, it's just that ChatGPT feels more flexible.

After years of treating faster, bigger, and more capable AI as something approaching a moral imperative, the people running some of the world’s most powerful AI companies suddenly agree that perhaps everyone should ease off the accelerator.

Anthropic CEO Dario Amodei kicked off the latest round with an essay titled “We Must Pace the Frontier,” warning that AI capabilities are advancing faster than the safeguards needed to contain their risks. OpenAI CEO Sam Altman and Google DeepMind CEO Demis Hassabis quickly endorsed the idea. It's extraordinary that the leaders of companies locked in one of the most expensive technological races in history are publicly agreeing that the race itself needs to slow down. One awkward detail is buried beneath the sudden outbreak of corporate caution. Namely that nobody involved has said which forthcoming frontier model will arrive later because of it.

“Pacing the frontier” can mean almost anything until somebody attaches a calendar to it. None of the companies supporting Anthropic’s proposal has publicly identified a forthcoming model it intends to delay as part of the initiative, nor has anyone defined whether slowing down means an extra week of safety testing, six months between major generations, or something more dramatic. For now, the most concrete commitments concern independent evaluators, monitoring, safety standards and coordination rather than an announced reduction in the cadence of frontier-model releases

It's odd enough to stand out even to those outside the tech space. U.S. vice president JD Vance said he felt “a little bit weird" about the fact that you have so many frontier AI tech companies kind of coming to the government and begging the government to regulate them, and that it came off as “a bit of a Trojan horse.” His skepticism does not settle whether regulation is necessary, but it highlights the contradiction running through the current debate.

Altruism or exclusion?

There are good reasons for the sudden anxiety. Amodei has warned about AI enabling cyberattacks, bioterrorism, economic disruption, and eventually systems humans could struggle to control. His latest proposal calls for independent safety evaluators with deep access inside frontier labs, coordination among companies on shared standards, and international cooperation around particularly dangerous capabilities.

That's rather different from the plain-English meaning of slowing development. Anthropic’s own recent history is ambiguous, as it paused external cyber evaluations of prerelease models and briefly stopped internal ones. It also paused higher-risk reinforcement-learning environments for several weeks.

Meanwhile, the frontier has continued moving. Anthropic released Claude Fable 5.1 and Mythos 5.1 this month, while OpenAI debuted GPT-6 Astra. Those launches preceded Amodei’s latest public call for an industry slowdown, but show the strange starting point for this new era of restraint. The companies asking everyone to discuss slowing down have just spent the month pushing the frontier forward.

If the companies genuinely believe development is moving dangerously fast, they already control their own research schedules and release calendars, while government regulation raises a separate question about whether rules designed with the biggest labs could also make life harder for smaller competitors.

Apocalyptic distraction

There is another problem with all this talk of existential danger. The more Silicon Valley discusses hypothetical superintelligence destroying humanity, the easier it becomes to overlook the considerably less cinematic ways AI is already hurting people.

Karolis Kaciulis, Lead System Engineer at consumer cybersecurity company Surfshark, argues that warnings about AI threatening humanity can amount to a “marketing move” that distracts attention from existing harms. “The threat itself is fictional, closer to a Skynet-style sci-fi scenario than the problems generative AI is already causing today, from automated scams to intimidation,” he said.

That skepticism deserves space alongside the warnings from AI executives. Generative AI is already making phishing, deepfake fraud, and automated scams cheaper and easier to scale. Questions remain about privacy, how information submitted to chatbots is handled, and the environmental cost of the infrastructure required to run increasingly large AI systems.

There is also a legitimate debate about how much progress the industry’s endless procession of model releases actually represents. Benchmark numbers climb, but determining whether each new large language model represents a profound new capability is considerably harder than reading a launch-day chart.

AI companies simultaneously warning about AI's power while still pushing ahead and sidelining safety comes off as bizarre to the average person. Amodei’s proposal is new, and judging it entirely by whether a company delayed a model by four days would be unreasonable. But his ideas need an independent evaluation system to have any muscle. The next step needs to be measurable, or it's irrelevant.