Employees are quietly turning to the likes of ChatGPT, Claude, and Gemini to draft emails, crunch data, and polish presentations. Some may hail a new era of efficiency, but in this latest incarnation of Shadow IT, ‘Shadow AI’ usage is brewing a new governance headache for businesses. 

But both end users who are pasting client proposals into AI chatbots and tech leaders who are overseeing them need to take note of one of the thorniest challenges facing IT and security leaders today: Shadow AI and the risk it poses. 

The Rise of ‘Shadow AI’ 

‘Shadow IT’ is a term coined to describe hardware, software, cloud services or similar used by employees for work without the knowledge or approval of the central IT or security department.

Salesforce CTA and Founder & CEO of EzProtect, Matt Meyers, told SF Ben that if a company genuinely wants to prohibit AI use, it should enforce strict controls, such as disabling access to AI endpoints like ChatGPT, Claude, and Gemini. 

But instead of banning AI, companies should embrace it to remain competitive, while establishing clear policies regarding data privacy, Matt said.

He told us: “If you’re not on AI and you’re not adopting AI, you are going to be behind all your competitors. And that’s the truth because more and more companies, even in regulated industries, are starting to adopt it because it is an accelerator. 

“Okay, there is the risk. If you’re putting customer data in there, like sensitive customer data, there is that risk that they could save it and use it with the model, right?”

Matt said that ChatGPT, Claude, and others have checkboxes to opt in or out of sharing your data with the provider, and he doesn’t feel it’s as big a risk as some people make it out to be. 

While you shouldn’t be putting private information that could hurt other people into LLMs, Matt adds that, for most companies, “their stuff really isn’t as secret as they think it is”. 

He said: “Most things people are putting in there are just like a sales pitch deck, or things like that. Just from my personal opinion, I think people are kind of going a little bit overkill on being afraid of AI and what it’s doing.” 

He said that he is more worried about what sensitive data the AIS (Artificial Intelligence System) is sharing with the public. Companies should have a policy for secure AI usage and should train employees not to hand out information to an LLM, the same way they are trained not to give out PII over the phone, Matt says. 

But the other thing to consider is the level of concern. Matt said: “I am not condoning sending private PII-type data to the AIS, not at all. What I am saying is that your data, your PowerPoints, your normal business docs – unless it’s a very sensitive doc like a pharmaceutical company’s formula for making their medicine – what you’re sending is not that innovative, it’s not something that your competitor probably isn’t doing already.”

AI Companies Using Your Data?

‘Shadow AI’ is a subsection of ‘Shadow IT’. When you ask Claude to help you with an email or presentation, that’s shadow AI. So what’s the problem? 

Firstly, storing corporate data in personal cloud storage could increase the risk of a data leak. Personal accounts are rarely as secure as corporate equivalents and lack guardrails and governance imposed by corporate IT and information security teams. 

Secondly, there could be gaps in compliance: unapproved tools bypass strict regulatory standards, such as GDPR, for instance, putting the company at risk. 

Thirdly, a company’s IT team cannot secure or patch assets which they do not know exist. 

We asked developer turned architect Beech Horn about Shadow AI and what implications there would be from employees in regulated industries feeding sensitive data into the likes of ChatGPT and Gemini. Beech said that the bottom line is that if a company is not providing adequate AI tools for staff, “they’re bringing their own”. 

“This is not a cat-and-mouse game you can win with a stick over a carrot,” Beech told SF Ben. He also mentioned that a number of AI companies use their users’ data for model training. According to the Claude website, Anthropic will use your chats and coding sessions (including to improve their models) if:

1. You choose to allow them to use your chats and coding sessions to improve Claude
2. Your conversations are flagged for safety review 
3. By otherwise explicitly opting in to training

Your Incognito chats are not used to improve Claude, even if you have enabled Model Improvement in your Privacy Settings, the company says.

The Social Media Comparison

People my age will remember when social media first emerged. The internet felt like something of a Wild West where nothing really mattered, and leaving a comment on Facebook was an ephemeral thing that you didn’t really think about. 

Over time, we came to realize that we had something called a ‘digital footprint’, and things we said had a much more permanent nature. Be careful what you say online – it could haunt you later in life, we gradually learned.

The dawn of LLMs proves once again the old adage: we study history so we don’t make the same mistakes. What you tell an LLM might not be instantly available on all your friends’ devices the same way a tweet would, but that doesn’t mean you should treat AI tools with any less discretion than you would a public forum – especially when it comes to sensitive information. 

Matt agrees with the social media comparison, telling SF Ben: “I feel like it’s kind of like the same thing in some ways… You don’t post stuff that you shouldn’t. It’s like not emailing out a presentation with people’s government IDs in it to someone else. 

“I think there’s the whole idea of training your people to be responsible for it, but if you just shut it out completely, you’re going to be left behind by your competitors because that’s the way the world’s going.”

Conversely, AI companies themselves also need to act ethically and be able to disclose what they’re doing, says Matt. People who use their services should know how their data is being used, and customers should not automatically be opted in to sharing their information. 

He added that recent breaches of AI have been troubling him and might be a reason to be more careful about what information you submit to the LLM. 

Final Thoughts

Shadow AI is more of a people problem than a technology one. AI tools make employees’ lives easier. Business leaders should respond with education and enablement rather than blanket bans. 

As Matt says, businesses that don’t use AI at all are at risk of being left behind. “If you’re really that concerned, then do pay a lot of money to get contractual agreements about how your data will be used with them,” he said. 

The tools have changed, but the principles really haven’t. 

The Author