The trust gap: is AI detection software the answer to building workplace trust?
With a third of employees admitting to use of unsanctioned AI to fulfil their role, how will the development of detection software restructure trust in the workplace?
AI has become fundamental to the modern workplace, integrated seamlessly into many of our day-to-day lives. Whatever your personal feelings on the hot-topic, business’ have embraced its ability to cut cost, streamline content and automate tasks that used to sit heavily on the shoulders of the workforce.
A recent survey by Blanchard (2026) revealed how the use of unsanctioned, shadow AI is eroding at the essential trust long found between employer and employee. Nearly 43% of the leaders it surveyed reported to having observed ‘undesirable’ use of AI in the workplace; ranging from the quiet judgement associated with colleagues relying too heavily on its content, to the mismanagement of verification when it came to data pulled from LLMs.
Roughly 24% of their the participants recorded that these attributes had become somewhat commonplace, while a fractional 18% actually admitted to partaking in the behaviours themselves.
Most surprisingly however, respondents were 2.4% more likely to report a colleague’s misuse of AI, than confess to their own usage of it.
The casual utilisation of AI has led to its own complications, with filtration and factchecking surrounding ‘workslop’ amassing its own can of worms altogether. A Chief Data Officer report in conjunction with Deloitte (2026) revealed that 65% of employees are trusting their AI output, even when data gaps emerge. A further 76% reported that their employer’s governance of AI was simply not keeping up with its rising usage across the business.
The evolution of AI detection software
However, amongst the torrent of generated content, AI detection software is developing almost as rapidly as its counterpart. Pangram, a 24-person startup founded by Tesla and Google alum, aims to become the golden standard in the game. The company has raised eyebrows for a different approach to distinguishing content made, or run through, LLMs.
Unlike traditional models, Pangram takes the queried text, converts it into a sequence of numbers, and runs it through a neutral network trained on roughly a million documents: half written by people, half written by models such as ChatGPT, Claude and Gemini. A final classification then determines a verdict on the passage, spread across four confidence levels.
Its latest detection model, Pangram 4, recognises both homogenous and heterogenous mixed text, while still managing to keep its false-positive rate low, sitting at around 0.0041% (about 1 in 10,000), out-performing trained humans in tests.
It does however have its limitations. The model can be inconsistent when running partial content against a full paper or report; the additional context giving the programme more information to make a fully informed judgement. Even weighted against its shortcomings, Substack recently pioneered its software, allowing users to scan any content valuing over 100 words on their writing platform.
The risks associated with unsanctioned AI
Reuters (2026) found during a global study that more than a third of professionals were using unsanctioned AI for work, raising concerns against subsequent security implications. Steve Hasker, CEO of Thomson Reuters, noted recent instances of lawyers using AI without company approval, impacting court cases and further damaging their firms’ reputations in the process.
In February, U.S district judge Jed Rakoff ruled that any information shared with Claude would no longer be considered privileged in litigation, citing the chatbot as a third party in United States v. Heppner. With such, a larger discussion has materialised surrounding privacy law, data protection and cyber-security.
The trust between employer and employee
Trust has always been critically intertwined with engagement, consequently influencing productivity, performance and ultimately affecting revenue. Trusting employees were 260% more motivated to work, and 50% less likely to search for another job according to MIT’s Sloane Management Review (2023), highlighting its prominence, but notably importance.
As detection software becomes more competent, it’ll be interesting to watch its role evolve across the spectrum of industry. Particularly, whether its implementation helps, or hinders, the foundation of trust between employer and employee.
