Social media that still feels real

Open any feed for a minute and count the pictures that could not have happened. A product floating in perfect light, a city at an impossible hour, a face a little too smooth. AI has made content quick and cheap to produce, and for the first time a small brand can post every day without a shoot.

That is a real gain, and we use it. It also creates a new problem. When anything can be made, people start to ask what was real, and the answer changes how they feel about the brand that posted it.

Our view is simple. Use AI for what it does well, and keep the parts people come to social media for in human hands.

What audiences say

The surveys point the same way. In a Pew Research Center survey of 5'023 American adults in June 2025, 76 percent said it was extremely or very important to be able to tell whether pictures, videos and text were made by AI or by people. Just over half, 53 percent, were not confident they could tell.

Getty Images asked more than 30'000 adults in 25 countries and published the results in April 2024. Of those, 98 percent agreed that authentic images and videos are pivotal in establishing trust, and almost 90 percent wanted to know whether an image had been created with AI. Getty also found that people feel less favourably towards brands that use AI generated visuals to create people or products.

Sprout Social’s 2026 influencer marketing report, based on 2'250 consumers in the US, the UK and Australia, found that 44 percent are uncomfortable with brands using AI influencers. The same report found that audiences see frontline workers as significantly more authentic than polished brand accounts, and that 40 percent discover new products through content made by employees every month.

What the platforms now require

The platforms have turned the mood into rules. Meta labels AI content on Facebook, Instagram and Threads when it detects industry standard signals such as C2PA and IPTC metadata, or when the person posting discloses it. Since July 2024 the label reads “AI info”. Meta requires people to disclose photorealistic video and realistic sounding audio made with AI, and says it may apply penalties when they do not.

TikTok said in May 2024 that it had required creators to label realistic AI generated content for more than a year, and it began reading Content Credentials to label AI content made on other platforms automatically. YouTube asks creators to disclose realistic altered or synthetic content, may add a label itself, and shows a more prominent label on videos about health, news, elections or finance. In July 2025 it renamed its repetitious content policy to inauthentic content, to make clear that mass produced and repetitive videos cannot earn money, including AI content built from generic templates.

In Europe the law follows. Article 50 of the AI Act applies from 2 August 2026. Providers of generative systems have to mark their output in a machine readable way, with time until 2 December 2026 for systems already on the market, and anyone who publishes a deepfake has to disclose it.

The real touch

Faces come first. People follow people: a founder explaining a decision, an employee showing how something is made, a customer telling the story in their own words. These are the posts audiences trust, and they are the ones a model cannot fake without crossing the line the platforms and the law now draw.

Voices come next. A real voice carries hesitation, accent and humour. It is also easy to clone, which is exactly why it should stay real whenever the person is real.

Then moments and the work behind them. The event that happened, the team at the stand, the first delivery of a new product, the workshop, the draft on the wall, the mistake and the fix. They cost almost nothing to film, and AI cannot invent them without lying.

These are also the parts of social media that need judgement: what to show, what to leave out, when to answer and when to stay quiet. That judgement, and the responsibility for every post that goes out under a company’s name, stays with a person.

Where AI belongs

AI is very good at the rest. Product pictures in every format, in a style that stays the same from the first post to the hundredth. Backgrounds, variants and sizes for each channel. Captions and drafts in the brand’s voice, ready for someone to correct. More ideas than a team could write, and a first sorting of which ones have a chance. Reports that show what worked.

The line we draw is honest representation. A made picture of a real product is fine, labelled where the platform asks for it. A made picture of a person who does not exist, presented as a customer, is not. A cloned voice of the founder saying something the founder never said is not. If a viewer would feel deceived on finding out how a post was made, it should not be made that way.

The right mix

There is no published ratio that works for every brand, and we would distrust anyone who sold one. A useful start is to divide the feed by job. Content that sells a product can be largely made with AI, as long as the product is shown as it is. Content that builds trust, the people, the voices and the moments, should be real. Content in between, explainers, tips and announcements, can mix both: a real person on camera, with made visuals around them.

Then measure. Post both kinds, compare reach, saves, comments and the messages that follow, and let the numbers move the mix month by month. The labels are not the enemy. A clear label on a good made picture costs little. Being found out costs a great deal.

How we run it

Tempo runs social media this way. The strategy sets the mix of real and AI made content for each channel, the content engine makes and tests the month, and today a person approves every post. The visuals come from Reel, in the brand’s locked styles, and a person picks every image and every clip that is used.

What a company can do this month: list the people, places and moments only it has, and plan the calendar so every week carries at least one of them. Turn on the platform labels where they apply, keep the original files, and write down what you will never generate. Then let the results tell you whether the mix is right.