At first, one could tell that the other person was human by asking them to type the correct CAPTCHA or tick all the traffic lights in a 9×9 grid, but today, the internet is a much harder place to find out the truth. As AI-generated text, images, voices and videos flood our social media feeds and agentic AIs are running firms, even something as familiar as a cat batting her eyelids in a video can leave us wondering if that is actually a cat, or a very convincing digital feline.
Take AI writing tools, for instance. Give one a prompt to write an essay on the cow, and you could get a grammatically perfect write-up explaining the origins of bovines to an argument about how veganism could save the planet. The problem is that when millions of such sentences begin circulating online, the patterns become recognisable, making it paradoxically less human. That has created a new category of AI tools designed to make AI sound, look and behave more human. Text is the most obvious. Tools such as WriteHuman and StealthWriter take AI-generated drafts and rewrite them for natural phrasing, rhythm and variation. WriteHuman can also provide sentence-level AI probability analysis, while StealthWriter offers different levels of rewriting and sentence-by-sentence alternatives.
Refining Synthetic Interaction
The idea of ‘humanising’ text is increasingly about making machine-generated content more context-aware, with no repetition and more conversational, especially when it comes to agentic AI, where systems have been developed in a manner that can plan and do your work without being prompted constantly. Frameworks such as CrewAI allow developers to give agents distinct roles, while LangGraph supports workflows in which an agent can pause and bring a human into the loop. Microsoft’s Copilot Studio similarly allows organisations to customise how conversational agents respond and behave. They aim to make an AI agent more like a colleague who knows when to ask for follow-ups and when to act according to the given work.
Then there is the voice aspect of content. In 2026, text-to-speech has moved well beyond the robotic narration once associated with automated phone calls (even spam calls). For example, software company ElevenLabs generates speech with control over voice tone, pace and emotional delivery, while its voice library offers thousands of expressive voices. OpenAI’s voice systems are similarly designed around more natural, expressive interaction.
Hume AI takes the idea further with its Empathic Voice Interface. It analyses cues such as the rhythm, pitch and tone of a user’s speech and can alter its response accordingly, becoming more apologetic when it detects frustration or more sympathetic in response to sadness. The company calls them ’emotionally intelligent’ models. It is designed to make the media behave more like a chat between two people, including all the “umm” and “yeah, right”. Still, this is one aspect of humanising audio, as tools can also make imperfect recordings sound professionally ‘human’.
For instance, Adobe Podcast’s Enhance Speech uses AI to clean up voice recordings and reduce unwanted background noise, making an imperfect human recording sound as though it came from a soundproof studio.
Video Realism
Video is where the distinction becomes even harder to spot, specifically even harder for generations before us who rely on WhatsApp forwards for information.
LA-based AI company HeyGen can turn a script into a talking avatar with synchronised lip movements and expressions, while its video localisation tools translate existing videos into multiple languages while preserving the speaker’s voice, expressions and lip movements. “Character consistency is what separates a truly useful avatar from a mere gimmick. Avatar V delivers this across every angle, expression, and video you create,” the company says on website. It is built on a five-stage AI training pipeline that captures how you move, gesture, and express yourself.
The AI-human contest is becoming a two-way game. Recently, online publishing site Substack started letting readers scan posts, notes, comments and replies published from July 21, using Pangram, an AI text and image detector, to estimate how much text was human-written or AI-assisted. The result is a growing tech layer on both sides: tools that make AI content more human-like and tools that try to determine how much real human thought went into it.