Would my boss care if I bought ChatGPT Plus out of my own pocket to help me write an article?
This is the question that surrounds “shadow AI”which refers to using AI tools in the workplace without formal approval by companies.
Due to the pressure to work faster and the proliferation of easy to use generative AI tools for editorial staff, more are using AI in order to complete tasks. While using generative AI to perform minor tasks, such as grammar checks, rewriting of copy, or testing headlines, falls under one category of offenses, other uses could lead to bigger problems in the future if not checked. Legal experts warn that this could lead to inconsistent standards of editorial content, security vulnerabilities, and ethical lapses.
Let’s take a closer look at the issue and how publishers are positioned to handle it.
What is shadow AI?
Shadow AI is the use of AI at work which has not been approved or licensed. It’s been a Unauthorized use of AI-generated tools can be a thorn in the sides of IT departments around the world, as it can make businesses more susceptible to data breaches.
However, its application in newsrooms presents a unique set considerations. It could be illegal to enter sensitive data, such as sensitive source material, proprietary information, embargoed stories, and copyrighted materials, into large language models, without the publisher’s approval. This could compromise the protection of the information, the journalist’s work, and their reputation.
If someone takes [my work] into a system and the owner of that systems has a copy, they could use it in ways [I] would never have intended or allowed. Michael Yang, senior director for AI advisory services at Husch Blackwell, said that one of the ways to train AI is through a computer program. You could find yourself in a situation in which you have unintentionally or inadvertently caused a breach in contract.
What are the risks?
Legal professionals who spoke to Digiday mentioned three main concerns: the potential bias in AI models, confidentiality issues and accuracy questions.
The biased nature of AI models has long been reported. Yang said that if the data used to build AI models is biased (such as skewed towards certain races or genders), and journalists rely on tools created from these models, the output may perpetuate stereotypes.
Copyright infringement cases, such as the one brought against OpenAI by The New York Times, are based on LLMs scraping online content from publishers and using it to build their models. Felix Simon, Oxford University’s research fellow for AI and news who studies the implications for journalism, said that the same questions about how these LLMs use data to train their models are also why it could be dangerous for journalists to enter copyrighted information (or sensitive information, such as confidential source data) into an AI system that is connected to internet and not locally hosted.
Simon said that sensitive data could be fed to these unapproved systems, which would then be used to train the AI models. This data may appear in the outputs. He added that if these systems were not secure, they could have been viewed by the AI tech companies, people reviewing model outputs for updates to the system, or even third parties.
Yang said that sharing copyrighted data with an AI system in this manner could be illegal because AI companies can ingest and use inputs as training data. Gary Kibel, a partner at Davis+ Gilbert who advises clients in media and advertising, said that the publisher could be held liable for infringing copyright or creating infringing material.
Using a tool that isn’t vetted could cause accuracy issues. “If you enter into an AI platform ‘If CEO Jane Doe does the following, what will that mean?’ then the AI platform adds that to their training data and someone else’s output shows that CEO Jane Doe has done the following… they might come to you and ask, How did this get out?” Kibel said. Many of the larger publishers have established formal policies, principles, and guardrails to govern their newsrooms.
Gannett’s ” Ethical Guidelines for Gannett Journalists Regarding AI Generated or Assisted Content (19459019)” is one of a few publishers who have developed this policy — others include The Guardian The New York Times The Washington Post Publishers have also created internal groups to determine these principles and guidelines.
Gannett, for example, has an “AI Council” made up of managers from different departments who are responsible for reviewing and approving new AI tools and cases. In 2023, similar task forces were created at companies such as BuzzFeed and Forbes.
A spokesperson for the company stated that “These protocols ensure protection of Gannett personnel and assets, intellectual property, and information.”
It is important to educate newsroom staff about the risks of using AI tools that are not approved or paid for by their company. An anonymous publishing executive said that the best way to approach the issue is to explain to employees the risks, and to do so in a personal manner. Publishers believe that their AI-dedicated tasks forces, policies, and guidelines are enough to steer the newsrooms into the right direction.
This could work as long as the guidelines “have teeth” and that consequences, such as disciplinary actions, are clearly explained. Yang, a former director at Adobe, is an expert in this area.
Companies may also whitelist technology approved for use in the newsroom. The New York Times, for example, recently approved a number of AI programs, including Google’s Vertex AI, NotebookLM and other similar tools, for editorial and product staff. Semafor reported that
It’s difficult to do that if you are a small publisher. It’s impossible to review every AI tool that a journalist could use. Legal experts and publishing executives told Digiday that they understand the difficulty of controlling how journalists utilize online information.
How can you police shadow AI?
You can’t. Not completely. You can let your staff know how you and the company feel about AI tools at work. Yang said, “It may be as simple as a person running an app on a private phone.” “How can you police this when it’s their phone, their property, and they can do it with no one knowing?”
That’s where formal policies, principles, and guardrails established by publishers can be helpful.
A publishing executive said that they expected some shadow AI in the newsroom but were confident in the training provided by their company. The company holds several training sessions per year to discuss its AI policy and guidelines. For example, they do not allow employees to upload confidential material, financial data, or personal information to LLMs that have not been approved by the company.
The executive said, “I trust people to make judgments about the work they do and what’s best for them.”
According to a spokesperson for Gannett, the company has an “effective process” in place to approve and implement technology across its newsroom. The company has a tech policy that details the software and online services approved as well as how access to and payment for other services can be requested if necessary.
The spokesperson stated that “this policy helps us to ensure the integrity and security of our systems and data.” According to a report from AI software company Trint
64% of organizations intend to improve employee training and 57% plan to introduce new policies regarding AI usage this year. But companies should also ask themselves: Why are journalists doing this kind of thing?
The said, “Maybe the tools available to them aren’t sufficient.” “You can lean in and say, We’re going test the tools, have protections for them… and we’re gonna have policies and education so that you understand what you can do and what you can’t [and] and how to best use it to avoid these problems.