An honest starting point
It would be convenient to offer a checklist that reliably separates machine-made content from human-made content. No such checklist exists. Generative systems keep improving, people edit what they produce, and many human writers and photographers produce work that looks “machine-like” by any rule of thumb. This article therefore covers clues that sometimes help, explains why they are weak evidence, and suggests habits that work better than trying to judge by eye.
The broader category of text, images, audio and video made or heavily altered by software is often called synthetic media. The basics of how it is produced are covered in the pillar article on generative AI.
Clues in images
Earlier image generators left obvious marks, and some still do:
- Hands and fingers with the wrong number of digits or odd joints.
- Text in the picture that looks like letters but spells nothing.
- Mismatched details, such as earrings that differ, glasses with uneven frames, or buttons that do not line up.
- Strange backgrounds, such as objects that blend into each other or architecture that does not make structural sense.
- Overly smooth or glossy surfaces, as if everything has been retouched.
- Inconsistent lighting or shadows that point in different directions.
These flaws come from how the systems build pictures from learned appearance rather than understanding, as described in the article on image generators. But they are becoming less common, and a flawless image proves nothing. Equally, an odd-looking photo can be an ordinary photo taken with an unusual lens, heavy compression or a photo editor.
Clues in text
For writing, people often point to features such as:
- very even tone and a tidy, balanced structure
- generic phrasing and repeated sentence patterns
- confident statements with no specifics, sources or personal detail
- facts, quotes or references that cannot be found when checked
- a habit of ending with a neat summary
Some of these do appear often in unedited output. However, they also describe a lot of ordinary human writing, such as corporate memos, school essays and marketing copy. Non-native speakers and people who write in a plain, formal style are particularly likely to be wrongly judged as machines. A person can also prompt a tool to write in a messy, personal voice, or edit its output until the traces are gone. Anyone who has seen how prompt wording changes style will know that style is easy to steer.
The one clue that carries more weight is verifiability. A fabricated citation, a nonexistent book or a made-up quotation is a classic sign of AI hallucinations, and checking is straightforward. Even so, careless humans also make errors, so a mistake alone is not proof of a machine.
Why detection tools are imperfect
Software exists that claims to estimate whether text or an image was machine-made. It is worth treating such scores with caution:
- They give probabilities, not proof, and can produce both false positives (flagging human work) and false negatives (missing machine work).
- They are generally weaker on short passages, edited material and text in less common languages.
- As generators change, detectors trained on older output can fall behind.
- A score can be gamed by lightly rewording or re-saving a file.
Because of the risk of false accusations, a detector result should never be the sole basis for a serious decision, such as penalizing a student or an employee. The cost of wrongly labeling a real person’s work as fake can be considerable.
Watermarks and provenance labels
Some developers and camera makers are working on technical markers that record how a file was made. These include invisible watermarks added to generated output and signed metadata that records a file’s history. They show promise but have limits: not every tool uses them, metadata can be stripped when a file is re-uploaded, and absence of a label does not prove authenticity. Standards are still developing, and it is too early to assume they will be widely available or reliable.
Habits that work better than spotting flaws
Since looking for flaws is unreliable, it is more useful to ask different questions:
- Where did this come from? Look for the original source, an identifiable author or publisher, and whether others independently report it.
- Does it fit what else is known? Search for the claimed event or quote and see whether credible outlets confirm it.
- Can the image be traced? A reverse image search can show earlier appearances or the original context.
- Is it trying to provoke a reaction? Content built to trigger anger or alarm deserves extra scrutiny, particularly given how quickly false material travels, as discussed under disinformation spread.
- How much does it matter? A decorative illustration needs little checking. A claim about health, money or an election deserves much more.
For video and voice, the issues are somewhat different; see the article on deepfake technology.
Where this leaves us
You will not always be able to tell, and neither will experts or software. The practical response is to shift attention from “was this made by a machine?” to “is this reliable?” That means checking sources, being open about your own use of these tools where it matters, and being careful about accusing others. Disclosure norms and rules are still being worked out and differ by school, workplace and country, so check the expectations that apply to you.