Your CEO said one thing. Each investor may hear another.
A Microsoft filing describes live video tailored to every viewer, so boards must decide now which version of their words counts.
Sign up for my Substack daily AI newsletter: https://johanosteyn.substack.com/
Follow me on LinkedIn: https://www.linkedin.com/in/johanosteyn/
Picture your chief executive presenting annual results on a live webcast. One investor watches in English, another in isiZulu, a third through a filter that removes every hesitation. Each sees the same face, with lips moving in time with the words, yet each hears something slightly different. A recently published Microsoft patent application describes exactly that capability. It is not a product, and it may never become one. It does, however, raise a question South African boards have never needed to answer: when software can rewrite live speech for each viewer, which version of what your company said is the real one?
CONTEXT AND BACKGROUND
The application, published in the United States as US 2026/0341309 and filed by Microsoft Technology Licensing in April 2025, describes AI agents that find words in live audio and video, remove or replace them, and generate new frames so that the speaker’s lips match. It can draw on viewer profiles, including language, age, education and location. It describes keeping the unedited recording as “ground truth”, although that does not guarantee every implementation would, and it does not establish that viewers would be told anything had changed.
The building blocks are already in ordinary offices. Microsoft’s Teams Interpreter agent translates meetings in real time and can simulate each speaker’s own voice in another language, and Microsoft’s administrator guidance lists “Simulate my voice” as the organisation default. Human eyes offer little protection. In a 2025 study reported in August 2026, 25 of 34 students, or 74%, did not notice that a live video of their own professor had been altered with a freely available face-swapping tool.
INSIGHT AND ANALYSIS
Unlike a criminal deepfake, this involves a real speaker, a mainstream platform and an alteration that may be lawful, even helpful. What shifts is authority. A live broadcast has always been a shared record: everyone heard the same words, and the speaker answered for them. A viewer-side edit moves the decision about what was said from the person speaking to the settings of the person listening, while accountability stays exactly where it was.
For a listed company, that matters. A filter tuned for clarity could drop a forward-looking caveat or smooth over an executive’s awkward pause under questioning, and two shareholders could leave the same meeting with different understandings of the same statement. South Africans have already seen the criminal version. Maya Fisher-French recently warned that a video using her and Bruce Whitfield “completely fabricates a broadcast”, with their voices and footage synthetically altered. The legitimate version would be far more polished.
I have previously written about this in a related setting, arguing that the greater danger of synthetic media is how easily genuine evidence can be dismissed as fake. Personalised editing feeds that problem from both sides. An executive could disown words a viewer really heard, and the viewer could never prove which version was authentic.
The strongest counter-argument deserves weight. Many patent applications are defensive, dubbing and editing are old practices, and live translation across South Africa’s 11 official languages would be a real gain. Even the accessibility case cuts both ways, though. A recent viewpoint in a speech-pathology journal urges acceptance-based approaches over fluency goals, which makes an automatic “stutter removal” setting, whatever its intent, a judgement about whose speech counts as normal. The same logic applies to accents, and to the code-switching that makes South African English what it is.
IMPLICATIONS
Regulation is moving, unevenly. From 2 August 2026, the EU AI Act requires deepfakes to be clearly labelled. In South Africa, Minister in the Presidency Khumbudzo Ntshavheni told Parliament in May 2026 that the national AI policy should compel digital platforms to disclose AI-generated content. That policy is being redrafted, which leaves room to deal with altered live speech explicitly. I should declare that I serve on a working group in the national AI policy process.
Boards need not wait. First, record and retain your own unedited copy of every material public statement, from results presentations to annual general meetings, and treat it as a disclosure record. Second, decide which tools may alter your executives’ voices or faces, and check the defaults in the software you already license. Third, add clauses to speaker and media agreements that prohibit live alteration without consent and disclosure.
CLOSING TAKEAWAY
For most of history, seeing someone speak was proof that they had said it. That link is weakening, and not only because of criminals. When software decides what each viewer hears, the speaker’s words become a draft, and the audience’s settings become the editor. Companies cannot stop that technology from arriving. They can control the one thing that settles every dispute: an original record they hold themselves. Those that keep it will know exactly what they told the world. The rest will be left arguing with their own reflection.
Johan Steyn is a prominent AI thought leader, speaker, and author with a deep understanding of artificial intelligence’s impact on business and society. He is passionate about ethical AI development and its role in shaping a better future. Find out more about Johan’s work at https://www.aiforbusiness.net
Comments