In the field of generative artificial intelligence, there is a constant race to make interactions with models as natural as possible. As part of its new wave of models, OpenAI launched two versions on July 8, 2026: GPT-Live-1 and the lighter GPT-Live-1 mini. The goal is to remove the barrier between human and machine through instant, no-latency communication and the ability to actively listen. However, as a recent experiment by viral influencer known as Husk showed, "humanity" in voice does not necessarily mean higher intelligence in logical tasks.

Husk's test: When AI "breaks" on simple spelling

Influencer Husk, who became famous by repeatedly finding weak spots in ChatGPT models, put GPT-Live through his classic test. Instead of complex math problems, he opted for something any primary school student should handle: counting letters within words.

In previous versions of ChatGPT, Husk managed to make the model stumble on words like "December" or "strawberry." With the new GPT-Live model, set to its highest intelligence level, a significant improvement was expected. However, when Husk asked how many "e" letters are in the word "seventeen", the answer was disappointing. The model claimed there are only two "e" letters in the word (one in the "seven" part and one in "teen"), even though there are actually four.

This incident, described in Fast Company, points to a deep problem that plagues all large language models (LLMs), regardless of how good they sound.

Why can't AI see letters? Technical explanation of the problem

To understand why GPT-Live fails at tasks that are trivial for us, we need to understand the concept of tokenization. LLMs don't work with individual letters the way we do. Instead, they break down text into smaller pieces of information called tokens. A token can be a whole word, part of a word, or even a group of characters.

For the model, the word "seventeen" is likely one or two tokens. The model thus "sees" the meaning of the word and its context, but doesn't have direct access to its visual or structural composition from letters, unless explicitly provided by a decomposition process. This is a fundamental difference from traditional algorithms that would easily count spelling. For an LLM, a word is more of a mathematical vector of meaning than a sequence of characters.

Comparison with competitors: Gemini vs. Claude

OpenAI is not alone in this regard. Let's look at the current state of the market:

• Google Gemini (Live mode): Google is striving for a similar voice integration, where Gemini leverages strong ties to the Google Workspace ecosystem. Gemini often performs better in data integration, but in conversational naturalness it still competes with OpenAI.
• Anthropic Claude: Claude is known for its high level of logical accuracy and "safety-first" approach. Although Anthropic doesn't have as aggressive a voice interface as OpenAI, their models tend to hallucinate less in text-based tasks, even though they still suffer from similar tokenization limits.

Practical impact: What does this mean for users and businesses?

For the average user, this means that GPT-Live is a fantastic companion for brainstorming, learning foreign languages, or simple voice chat, but you must not rely on it for precise tasks requiring text work (e.g., real-time spell checking or code analysis).

For businesses, this gap between "human voice" and "logical accuracy" is critical. If you plan to implement AI voice assistants for customer support, you must bear in mind that the model can sound completely confident and human, while providing factually or spelling-wise incorrect information in the actual content. This can lead to erosion of customer trust.

Availability in the Czech Republic: OpenAI generally makes its services available globally. GPT-Live is accessible to users with a ChatGPT Plus subscription (approximately $20, which is roughly 475 CZK per month). Czech language support in voice models is constantly improving, but minor phonetic errors may still occur in real time, caused by the fact that models are primarily trained on English data.

In the context of the EU AI Act, it is important to monitor how these voice models handle biometric data. If the model "listens" and responds to emotions in the voice, it enters an area of regulation that requires a high degree of transparency and protection of user privacy within the European Union.