Capabilities and Limits
An honest map of what 2026 AI models genuinely nail — and where they still faceplant, from hallucinations to melted text.
To create great things with AI, you need two maps: one of what these tools do brilliantly, and one of where they quietly fail. Marketing gives you the first map. This lesson gives you both — because the difference between a frustrated beginner and a confident creator is knowing which battles the model will win *before* you spend tokens on them.
What 2026 models genuinely nail
- Photorealistic images that pass a casual scroll — lighting, skin, materials, reflections.
- Consistent artistic styles, from watercolor to isometric 3D to film noir.
- Short cinematic video with synchronized sound — Veo 3.1 and Sora 2 generate speech, ambience, and effects with the visuals.
- Complete produced songs: vocals, lyrics, arrangement, and mix in one pass with Suno.
- Speed: iterating ten ideas now costs minutes, not a design sprint.
Hallucinations: confident nonsense
A hallucination is output that is fluent, confident, and wrong. LLMs invent citations, dates, statistics, and API functions — not because they are "lying," but because they predict *likely* text, and a plausible-sounding fact is likely text. Remember How LLMs Work: prediction, not lookup. Image and video models hallucinate too, in their own dialect: melted logos, impossible reflections, staircases to nowhere.
The classic artifacts
- Text inside images — a short sign in quotes usually works on Imagen 3; a full paragraph still melts into alphabet soup.
- Hands and crowds — main subjects are mostly solved in 2026, but background people still grow creative anatomy.
- Video physics — in longer clips, objects can morph, teleport, or forget gravity exists.
- Faces of real people and brand logos — often blocked by safety filters, and legally risky anyway (see the ethics lesson).
Knowledge cutoffs: frozen in time
Every model learned from data collected up to a certain date — its knowledge cutoff. Ask about anything newer and it either admits ignorance or, worse, hallucinates an answer. Some chat products bolt on live web search to compensate, but generation models do not know last week's news, memes, or product launches. If your creative brief depends on a current event, you supply those details in the prompt.
| Limit | Practical workaround |
|---|---|
| Text in images melts | Keep it short and wrap it in "quotes"; regenerate just the sign with inpainting |
| One wrong detail in a great image | Fix only that region in Quick Edit instead of re-rolling everything |
| LLM-stated facts | Verify anything you would publish — every citation, date, and number |
| Same character across images | Use reference images — covered in the Image Generation path |
Getting text right in an image
A minimalist coffee shop storefront at dusk, a hanging wooden sign that reads "BREW" in bold sans-serif letters, warm window light, flat design illustration
Short text, in quotes, described physically (a hanging wooden sign). This works around the melting-text artifact instead of fighting it. Compare: asking for a full menu on a chalkboard — that is still a 2027 problem.
Got a great image with one wrong detail? Fix just that part, keep everything else. Fix it in Quick Edit