Is Spellcheck AI?

Defining AI & why it matters

“What is AI?”

I have been asked this question through the course of my research and I don’t like answering it. In the popular conscience there is a notion that AI must be as good as humans in their assigned task, otherwise it is just a regression- or rule-based piece of software. ChatGPT fills this requirement for most people; talking to it is almost like talking to a human, albeit a human that confidently makes up answers. When asking someone if they are using AI in their organization, they will have ChatGPT or RPA (Robotic Process Automation) in mind, not their company’s sales forecast or their Word spellcheck add-in.

In the legal world, however, there are several definitions that are converging on a common understanding of AI that it is in our popular understanding.

The EU AI Act (in the April 2021 text) defines AI as: “‘artificial intelligence system’ (AI system) means software that is developed with one or more of the techniques and approaches listed in Annex I and can, for a given set of human-defined objectives, generate outputs such as content, predictions, recommendations, or decisions influencing the environments they interact with,” where Annex 1 refers to “(a)  Machine learning approaches, including supervised, unsupervised and reinforcement learning, using a wide variety of methods including deep learning; (b)  Logic- and knowledge-based approaches, including knowledge representation, inductive (logic) programming, knowledge bases, inference and deductive engines, (symbolic) reasoning and expert systems; (c)  Statistical approaches, Bayesian estimation, search and optimization methods.” This would encapsulate most older decision-making or recommender systems, including regressions any human-programmed rule-based systems.

This is not the final definition of AI for the purposes of the EU AI Act; it has been amended to be closer to the OECD definition. The OECD states that “An AI system is a machine-based system that can, for a given set of human-defined objectives, make predictions, recommendations, or decisions influencing real or virtual environments. AI systems are designed to operate with varying levels of autonomy.” While not substantially different, this definition does not specify the techniques behind the models.

The National Institute of Standards and Technologies (US) will likely be setting the definition of AI for the US and derives their definition from the OECD and ISO/IEC: “The AI RMF refers to an AI system as an engineered or machine-based system that can, for a given set of objectives, generate outputs such as predictions, recommendations, or decisions influencing real or virtual environments. AI systems are designed to operate with varying levels of autonomy.” This definition is strikingly similar to the OECD definition, with the differences in emphasis, except for the omission of “human-defined” to describe the objectives.

Is spellcheck AI…?

Almost definitely. In the case of the standard, comes-with-Word spellcheck, we are talking about a machine-based software add-in that makes predictions and/or recommendations based on an algorithm, presumably rules-based.

The reason why I find this so compelling as a case is that universities, and presumably educational institutions more broadly, are looking at ways to regulate the use of AI writing aides by students and pupils in light of ChatGPT proving itself very helpful. They have several theoretical ways of going about this: outright bans or requiring disclosure. Outright bans would be technically challenging to enforce, as I have yet to see a tool that can distinguish AI-written work accurately enough to meaningfully support a ban. Disclosure would support students’ use of ChatGPT and other tools, while still letting faculty know which AI was assisting the student. A student could always fail to properly disclose, however, because of the lack of detection tools currently.

The bigger issue is that the definition of AI in the popular consciousness draws a fuzzy line somewhere between ChatGPT and the Word spellcheck you have been using for 20+ years. While universities will expect students to know what they mean, if they use generic language without defining what they consider to be AI they might be opening a can of worms when it comes to disclosure or bans, as the lack of specificity could lead to alternative interpretations.

To bring some examples, the University of Washington mentions AI content generator, specifically ChatGPT, within the context of its preexisting policies. While UW does not take a stance on whether AI content generators are a clear violation of any policy, they do remind students to keep existing policies in mind as they complete their coursework. Similarly, the University of York states that students would be at risk of academic misconduct if caught using generative AI in assessments or online exams. The existing policies are clarified in light of generative AI for these two universities without creating new policies and defining what AI is.

On the other hand, the Boston University’s Faculty of Computing & Data Science leans heavily into disclosure, while banning “AI tools” unless explicit permission is given. “AI tools” is explained to be “AI text generators and other AI-based assistive resources”. This mistake leaves the door open to students having to decide whether spellcheck and Grammarly (which markets itself as AI-powered!) are included in the disclosure rules and ban. The Hertie School makes a similar mistake in their discussion of “AI tools”, where they start off with precise language (AI content generation tools) but lead into undefined terms in their policy that 1) demands disclosure, 2) allow instructors to set the policy in terms of AI use in their course. These two universities attempt to make new policy, but make the mistake of using imprecise language that could be interpreted as banning or requiring disclosure of simple tools like spellcheck.

What should your organization do?

Organizations should have a clear working definition of what AI is. The OECD definition is a likely candidate to influence both the US and the EU in their eventual regulation, but this definition could change. A first step would be to start a catalogue of anything that could fall under the definition of the OECD definition of AI. Each AI-based system should be analyzed for the level of risk (unacceptable risk, high risk, low risk, or minimal/no risk). Each risk level has its own requirements: unacceptable-risk AI will no longer be allowed, low-risk AI will have transparency requirements, and no-risk AI will have no restrictions. If your organization is using high-risk AI, you will want to assess if the AI is worth continuing use. Using high-risk AI under the EU AI Act will require conformity assessments, registration in an EU database, disclosing accidents, and increased technical scrutiny.

While it is not sure how this regulation will be enforced, it would be wise to assume that any computer software or add-on might be AI and thus under risk of regulation. Although it is unlikely for the EU to interrupt your use of spellcheck, it will regulate it as part of the EU AI Act. Spellcheck would easily be considered a no-risk AI and no obligations would exist to restrict its usage.

P.S., ChatGPT agrees:

Yes, spellcheck is a type of AI (artificial intelligence) technology. Spellcheck software uses algorithms and rules to compare words in a document or text input to a dictionary of correctly spelled words, and then suggests corrections for any spelling errors that are detected. Some more advanced spellcheck programs may also use machine learning techniques to improve their accuracy over time, by learning from patterns in the types of mistakes that users make and adjusting their suggestions accordingly.