“Wrong” is doing a lot of work in that headline, and that is the first problem to untangle. When people say Copilot’s image generation in Word is producing wrong results, they almost always mean one of three completely different things: it refuses to generate anything at all, it returns an image that ignores half the prompt, or it throws a content-policy block on a request that seems entirely innocent. Each has a different cause and a different fix, and lumping them together is why so much of the advice online goes nowhere.

Here is the blunt version up front. As of mid-2026, the most common reason Copilot image generation in Word “doesn’t work” is not your prompt. It is licensing, usage caps, or a network setting silently breaking the request. The genuine prompt-quality problems are real, but they are the minority, and they are the easiest to fix once the plumbing underneath is actually working.

Problem one: it produces nothing at all

If you type a detailed image request and Copilot accepts it, says it has started, and then never delivers — or spins for thirty minutes and gives up — you are not looking at a prompting problem. You are looking at a broken pipe somewhere between Word and Microsoft’s image service. Work through these in order of likelihood:

Your licence. Since 15 April 2026, in-app Copilot features have required a paid Microsoft 365 Copilot licence. If image generation in Word stopped working in spring 2026 and never returned, this is almost certainly why. A regular Microsoft 365 subscription plus the consumer Copilot Pro plan is not always the same entitlement as the commercial in-app Copilot, and the messaging around this is genuinely confusing — people have paid for two different Copilot products and still found the in-Word feature greyed out.

Usage caps. Microsoft applies daily image-generation limits and dynamic throttling, especially on Personal and Family plans where each generation draws down a monthly AI-credit allowance, and on the free commercial Copilot Chat tier, which has an undisclosed daily cap. The tell-tale sign is that it worked fine yesterday and today reports that image generation “isn’t available for your plan.” That phrasing makes it sound permanent. It usually is not — it resets. If you burned through a batch of images in one session, you have probably hit the ceiling for the day.

Network blockers. This one is underrated and catches even home users. Copilot’s image generation runs through a web component, and VPNs, corporate security proxies, and aggressive “security filtering” features routinely break the exact calls it depends on. If nothing renders, turn off the VPN and retest before you assume anything is wrong with Office itself.

Add-ins. Add-ins can interfere with the webview Copilot uses. Run Word in Safe Mode and try again; if generation works there, disable add-ins one at a time to find the offender.

Isolate the service. The fastest diagnostic of all: go to Copilot on the web and try to generate the same image there. If it fails on the web too, the problem is your account or a broader service outage, not Word. If it works on the web but not in Word, the problem is local — channel, add-ins, or the install. This one test saves you from the Office-reinstall rabbit hole that Microsoft support so often sends people down, frequently with no result.

While you are checking, confirm connected experiences are on: File → Account → Account Privacy → Manage Settings, with both “Experiences that analyse your content” and “All connected experiences” enabled. With these off, the feature has no path to the service. The same plumbing underpins Copilot Design Suggestions in PowerPoint, so if you fix it for one app, you have likely fixed it for both.

Problem two: the content-policy block

This is the one that makes people genuinely angry, because the request seems harmless and the refusal feels arbitrary. You ask for something ordinary, and Copilot tells you a Microsoft policy has been triggered and to reword and try again. You reword it a dozen ways and still hit the wall.

The honest situation: Microsoft’s image-safety filters are deliberately conservative and sometimes wrong. They over-block. Following well-documented misuse of these tools, Microsoft tightened the guardrails considerably, and the filters now catch a meaningful amount of safe content as collateral. That is a defensible trade-off on Microsoft’s part, but it does not make it less frustrating when you are trying to illustrate a perfectly reasonable document.

What actually helps:

  • Strip anything the filter could misread. References to real people, public figures, brands, anything that could be read as violent or political, or anything involving children tend to trip the filter even in benign contexts. Describe the scene and the style generically rather than naming specifics.
  • Reword toward the abstract. “A person looking stressed at a desk” clears more reliably than emotionally loaded phrasing. The filter responds to surface words more than intent.
  • Do not spam retries. Hammering the same blocked prompt repeatedly can itself trip a rate limit, after which everything fails regardless of content. If you hit a block, change the wording meaningfully before trying again.

If a prompt is genuinely being blocked and you cannot find a wording that clears it, that is a filter decision you cannot override from inside Word. The realistic move is to generate the image with a different tool and insert it, rather than fighting a filter that is working as designed even when it is wrong.

Problem three: it ignores your prompt

Now we are in genuine prompting territory — the image generates fine, it just is not what you asked for. Faces are off, the composition ignores your instructions, text in the image is garbled, the style is generic. This is the area where your input actually matters, and a few patterns reliably improve results:

  • Lead with the subject, then style, then composition. Models weight the front of the prompt. “A flat-design illustration of a wind turbine on a green hillside, minimal, corporate colour palette” beats a rambling sentence that buries the subject.
  • Do not ask for legible text. Image models are still poor at rendering words. If you need a label, add it in Word afterward rather than asking the model to bake it in.
  • Use the model picker if you have it. Recent builds let you choose the underlying image model. Different models have different strengths; if one keeps producing the wrong aesthetic, switch and regenerate rather than re-prompting the same one endlessly.
  • Iterate in small steps. Change one variable per attempt. Rewriting the whole prompt every time tells you nothing about what is actually moving the result.

A realistic expectation matters here too: in-document generation is convenient, not best-in-class. For genuinely demanding imagery you will get better output from a dedicated tool. Copilot’s value in Word is speed and staying in one place, not winning on fidelity.

A concrete before-and-after

To make the prompting advice less abstract, here is the kind of rewrite that actually moves results. A weak prompt: “Make me a nice professional image for my report about our company’s growth this year that shows we are doing well and growing fast in a modern way.” That is all intent and no specification — the model has to guess the subject, the style, and the composition, so it produces something generic.

A stronger version of the same request: “A clean flat-design illustration of an upward trending line graph, minimal, blue and grey corporate palette, plenty of white space, no text.” Subject first (the graph), then style (flat, minimal), then palette, then an explicit instruction to avoid the thing the model does badly (text). Same goal, dramatically better odds of a usable result on the first try. The lesson generalises: describe what you want to see, not what you want it to mean.

The Mac and the “my colleague has it” problems

Two situations deserve their own mention because they generate a lot of fruitless troubleshooting.

On Mac, image generation in Word has had a rougher ride than on Windows, with recurring reports of requests that are accepted and then never deliver. Microsoft’s moderators routinely ask for the exact Word and macOS build because behaviour genuinely differs between versions — which is a polite way of saying some builds are simply broken for this feature and a reinstall will not fix them. If you are on Mac and generation hangs indefinitely, do the web test first (it almost always works on the web), and treat the in-Word failure as a known build issue to wait out or report rather than a configuration you can repair.

The “a colleague has it and I don’t” problem is rollout, not fault. Microsoft ships these features in stages across tenants and regions, and image-related capabilities have lagged in some markets by weeks or months. If someone two desks over can generate images in Word and you cannot, and your licence and settings match theirs, the unglamorous explanation is usually that the feature has not reached your account yet. There is no toggle to force it forward.

A note on currency

This is the most volatile feature we cover, and we will not pretend otherwise. The licence tiers, the daily caps, the available models, and the regional rollout all change on a rolling basis — some markets get features weeks or months after others, so a capability a colleague has may simply not have reached your tenant yet. We re-verify this page against current builds and date it at the top. If something here does not match what you see, Microsoft has very likely shipped a change since our last check.

If your problem is specifically in PowerPoint rather than Word, the causes overlap heavily but the interface differs — see Copilot in PowerPoint: Designer pane image suggestions failing. For the suite-wide overview of where and why Copilot image generation breaks, see Microsoft 365 Copilot: image generation failures in Word and PowerPoint. And if you are wondering why Outlook does not offer the same image generation at all, that is deliberate, and we explain the gap in Copilot in Outlook: AI image features and what’s currently broken.