Your guide to AI and this issue

by Isabella Aranda Garcia & Gia Miller

Design by Anne Kennedy

>Welcome to our AI issue—an issue written and designed by humans about AI. If you’re new to the world of AI, a casual user or have a basic understanding but get lost when talking to “tech geeks,” this guide will help you better understand AI. We recommend you bookmark or dog-ear this page and return to it if you’re confused at any point while reading this issue. We can’t promise it will have all the answers, but it will certainly help.

What is AI?

>AI is the abbreviation for “artificial intelligence,” but it cannot do what a human can do. It cannot adapt to new and different situations or circumstances like we can. It does not have beliefs or feel emotions. It is unable to make ambiguous decisions with incomplete data, demonstrate novel judgement or understand truth. AI uses patterns to calculate what response, action or piece of text should come next, based on the information it received. But when used properly, it can improve your daily life.

[BEGINNER]

>AI is software that can analyze information and produce an output, such as a prediction, recommendation, decision or response. AI is likely a part of your everyday life; it filters your emails to remove spam, recommends movies on platforms like Netflix or music on Spotify, recognizes faces in your digital photos and provides directions while driving.

There are four broad categories of AI:

1. Recommendation AI

Suggests what to watch, buy or listen to.

2. Generative AI

Creates text, images, music, videos and code.

3. Recognition AI

Understands speech, images, handwriting and faces.

4. Predictive AI

Forecasts outcomes and helps make decisions.

[INTERMEDIATE]

>AI systems produce their outputs by training on large amounts of data rather than following manually written rules like a computer would. During training, a model is repeatedly shown examples (past purchases, labeled images, sample text) so it can learn and adjust, improving its ability to recognize the patterns tied to a specific task. This training process is what powers each category.

1. Recommendation AI

Learns patterns in user behavior.

2. Generative AI

Learns patterns in how text, images or code are structured.

3. Recognition AI

Learns patterns in speech or visual data.

4. Predictive AI

Learns patterns in historical outcomes.

Once trained, the model applies these learned patterns to new data it has not seen before.

[ADVANCED INTERMEDIATE]

> AI does not “understand” information like humans do. It uses the patterns it learned during training to infer what should come next or what output is most likely to fit the task. Although you probably won’t see it happen, a chatbot, for example, generates its response one small piece at a time by predicting what comes next—whether it’s a word or action—based on the prompt and its prior experience. This is why AI can write prose, recognize images or make useful predictions. However, it is also why AI appears confident when providing false information or repeating bias.

Who are the biggest players?

[BEGINNER]

Anthropic (Claude)

Founded in 2021, Anthropic centers its identity on AI safety. It’s best known for its Claude model family and has a strong developer and enterprise following. Claude’s “constitutional AI” guides model behavior with a written set of principles rather than relying solely on human feedback. It’s often praised for its exceptional coding performance as well as its writing and editing.

Apple

Apple’s biggest AI strengths are its hardware-software integration, custom chips and privacy-oriented on-device processing. Instead of a standalone model, you’re more likely to find Apple’s AI through features (e.g. autocorrect, voice dictation or Siri) that become native parts of your devices (iPhone, iPad, Mac, etc.).

Google DeepMind (Gemini)

Google AI is everywhere: Search, Android, Chrome, Workspace, Gemma, Gemini and Gemini Notebook. It’s also a rare full-stack player, designing its own TPU chips and running its own data centers. Gemini stands out for its ground-up native multimodality, which helps it understand and reason across text, high-resolution images, audio and full video simultaneously, without relying on secondary tool conversions or text-only transcriptions.

Meta (Llama)

Meta is the leading force in the open-weight model movement, using its Llama family to shape what developers, startups and enterprises can build without relying entirely on closed APIs. It boasts a massive open-source community and enjoys tremendous consumer reach through Facebook, Instagram, WhatsApp and Messenger.

Mistral AI

Mistral, Europe’s premier AI lab (based in Paris), was founded in 2023 and is known for its efficient language models, open-weight releases (models you can run and customize) and enterprise-oriented AI offerings.

Fun fact: Mistral’s founders include alumni of Google DeepMind and Meta.

OpenAI (ChatGPT, DALL-E, Sora)

OpenAI began as a nonprofit with a $1 billion pledge in 2015; today, it operates with a nonprofit parent and a for-profit subsidiary. OpenAI introduced ChatGPT and generative AI to the world in late 2022, sparking today’s AI boom. ChatGPT’s biggest strength is its versatility, and it’s particularly strong at reasoning through complex problems, writing and rewriting, coding, research, summarizing, brainstorming and working through tasks iteratively.

Perplexity

Founded in 2022, Perplexity’s clearest advantage is research; it’s considered an AI-native answer engine that researches the web in real time, synthesizes a direct response and provides inline source citations. It’s great for research, checking claims and finding sources. Professionals ranging from journalists and researchers to attorneys and data analysts turn to Perplexity for its multi-step searches and source-backed reports.

xAI (Grok)

Founded in 2023, xAI is Elon Musk’s AI venture, most well known for the Grok model series. xAI’s biggest strengths are its fast-moving model development and built-in distribution through the social media platform X. Grok provides real-time search, reasoning and coding, along with voice, image and video generation. But what sets it apart is tone; Grok was designed with “a bit of wit” and “a rebellious streak.”

[INTERMEDIATE]

Adobe

Adobe embedded Firefly—its generative AI product family—directly into software and hardware products people already use (Photoshop, Illustrator, Premiere, etc.). Firefly features controls that let users refine or edit work rather than simply generate a one-off image. Its models are designed for commercial use, trained on licensed Adobe Stock material and public-domain content, not on subscribers’ personal or generated content.

ChatGPT Work

An AI agent within ChatGPT that can carry out multistep work across approved tools and files as well as on your desktop’s (when permitted) local apps and a browser. It can research and analyze information to create or edit documents, presentations, spreadsheets, reports, PDFs, websites and other deliverables. You can follow its progress, answer questions, change direction and approve important actions during the process.

Microsoft

One of AI’s primary enterprise-distribution engines, Microsoft places generative AI directly inside products workers already use, including Azure, Microsoft 365 Copilot, GitHub Copilot, Windows and its developer tools. It turns model capability into institutional adoption, governance, security and recurring enterprise workflows.

Perplexity Computer

A general-purpose AI agent, Perplexity Computer can operate a computer to complete multistep tasks. Given a goal, it reasons through an approach, delegates parts of the work to specialized AI models and takes action. It can browse the web, use connected apps and tools and write code as well as produce research reports, documents and data analyses. It can also run continuously in the background, so the agent can work on tasks around the clock and be checked on or redirected from anywhere.

[ADVANCED INTERMEDIATE]

Amazon

Amazon operates one of the world’s largest cloud platforms and is a major financial and infrastructure partner of Anthropic. Amazon Web Services (AWS) offers multiple model choices through Bedrock while building its own AI chips, creating an alternative to dependence on Nvidia hardware and a single model provider.

AMD

AMD makes powerful chips for training and running AI models, as well as the software needed to use them. Its main AI products include Instinct data-center GPUs, EPYC server processors and ROCm, an open software platform that developers use to run AI workloads on AMD systems. It’s Nvidia’s main competitor.

Copilot Studio Computer Use

This business tool lets organizations build AI agents capable of working through websites and Windows desktop applications. After describing a task, the agent can visually interpret the screen and use a virtual mouse and keyboard to click buttons, navigate menus, enter information and complete steps in systems. It can extract information, process invoices, enter data or fill out forms.

Hugging Face

Founded in 2016 as a playful iOS chatbot app for teenagers, Hugging Face transitioned its purpose and is now an open machine-learning platform where developers and researchers share the models, datasets and tools used to build AI applications. Its Hub hosts more than two million models, 1.5 million datasets and 1.5 million AI apps.

Nvidia

Nvidia is the leading hardware engine powering the AI boom, designing the graphics processing units (GPUs) essential for training and running large models. It has a near-monopoly on high-end AI chips, and its GPUs and computing platform/programming model are the backbone that nearly the entire industry trains and runs on.

Terms you should know

[BEGINNER]

Algorithm 

A set of instructions or rules a computer follows to solve a problem or make decisions. AI systems use algorithms to analyze information and recognize patterns.

Bias 

When an AI system’s results routinely favor or disadvantage certain people or groups because of how it was designed or trained.

Chatbot 

An AI program that interacts with users through conversation (text or voice), answering questions or helping complete tasks. Examples include LLMs like ChatGPT, Claude and Gemini, and they’re also found on websites like Amazon, IKEA and Sephora.

Generative AI 

AI that creates new content—including text images, music, videos or computer code—based on patterns it learned from existing data.

Hallucination 

When an AI system confidently produces information that is wrong, made‑up or not supported by reliable sources.

Large Language Model (LLM) 

The technology behind AI chatbots like ChatGPT, Claude and Gemini. An LLM is trained on enormous amounts of text and predicts the next word in a sequence, allowing it to generate human-like responses.

Machine Learning 

Building AI systems that learn patterns from data and experience, rather than being explicitly programmed for every task. They can improve at tasks without being programmed for every rule.

Prompt 

The instruction, question or request you give an AI system to tell it what you want it to do. Better prompts generally produce better results.

[INTERMEDIATE]

AI Agent  An AI agent is an AI‑powered assistant that can pursue a goal for you, not just answer a single question. It usually runs on an LLM but adds planning, memory and the ability to use tools like Word, Excel, email, calendars, web browsers and other apps or online services. “Building” or “configuring” an agent means that you’re giving the AI a mission that includes a goal, rules and a format. Instead of stopping after one reply, an agent can take a goal you set—research a topic and format the findings in a specific way, manage a complicated schedule, plan a trip or event, or analyze data and create side-by-side comparisons—and then break the work into steps, use the right tools (such as search engines, files, apps or databases), check its progress and keep going until it’s done or needs your input.

Computer Vision 

AI that enables computers to recognize and interpret images and videos, such as faces or objects.

Deep Learning 

A type of machine learning that uses networks modeled loosely after the human brain to recognize complex patterns. It’s especially useful for image recognition, speech recognition and generative AI.

Deepfake 

An AI-generated image, audio recording or video that realistically imitates a real person.

Embedding

A method AI uses to convert information (like a word or image) into numbers so the system can understand how things are related in meaning. For example, “cat” and “kitten” would be embedded close to each other, while “cat” and “boat” would not. It creates relevant answers even if we don’t use the exact words or phrases.

Natural Language Processing (NLP) 

The branch of AI focused on understanding, interpreting and generating human language.

Predictive AI 

AI that analyzes data to forecast future outcomes or identify likely events, such as predicting weather, detecting fraud or estimating when equipment may fail.

Prompt Engineering 

The practice of designing effective prompts and instructions so AI systems produce more accurate and reliable responses.

Retrieval‑augmented Generation (RAG) 

A technique where an AI model first looks up relevant information (like from documents or a database) and then uses that material to craft a more accurate response. As an example, if you asked your company’s chatbot about their vacation policy, it would pull the information from your employee handbook instead of guessing from what it finds on the internet.

Token

Instead of reading entire sentences, AI breaks text into small pieces called tokens. A token could be a whole word, part of a word or even punctuation. Whenever you ask a question or receive a response, the AI uses tokens. Businesses and developers often pay for AI services based on the number of tokens an AI processes, and you’ll likely use tokens when deploying agents.

[ADVANCED INTERMEDIATE]

Context Window

The maximum amount of text or data an LLM can “see” at once when generating a response. If it has a larger window, it can handle longer documents or conversations.

Multimodal Model

An AI system that can understand and generate across more than one type of data—text, images, audio or video—within a single model.

RLHF (reinforcement learning from human feedback)

A training method where humans rate or compare model outputs, and the model is adjusted to create answers that are better suited to human preferences and safety standards.

Synthetic Data

Artificially generated data used to train or test AI systems. These are often produced by other models and designed to look like real‑world data. For example, a hospital might test a new AI scheduling assistant with fake records before exposing it to real patient records.

Transformer 

The neural network architecture behind most modern LLMs. It’s designed to pay attention to the relationships among all tokens in a sequence at once, improving context handling. So if you ask ChatGPT to rewrite a paragraph, it’s able to keep track of every word and sentence (tokens) at once instead of one at a time.

Vector Database

A database designed to store and search information (embeddings), so the AI can quickly find information that is similar to the request. For example, a boutique builds a searchable vector database of all its customers’ purchases, converting the product description into embeddings. When someone comes in to buy a gift, the employee can type “current products similar to what Michelle bought before” and receive a list of current items that are similar to what the customer has purchased in the past, even if it’s not the same brand or category.

Red-teaming (for AI) 

A systematic testing of an AI system where a human attempts to make the system fail or produce unsafe outputs before it is launched. This is typically done to discover vulnerabilities and improve safety.

Prompting Basics

If you’re getting generic or brief responses when you ask AI for help with an email or any other task, the problem is probably your prompt. AI is a brilliant intern, but it’s not a mind reader. Vague instructions lead to vague results. To unlock its true potential, you need to transform your simple requests into strategic directions.

[BEGINNER]

Let AI help you write the prompt

Not sure how to phrase your request to achieve the best results? AI can help. Tell your AI tool what you’re trying to accomplish in plain language, including the problem, your goals and your target audience. Then ask it to draft a strong and thorough prompt, instructing the AI to ask you any additional questions it needs to write the prompt.

Try: Act as a prompt design expert. Please create a reusable prompt that takes a long-form whitepaper and extracts a month’s worth of social media content. Please ask me any additional questions you need to write a thorough and complete prompt.

Then, copy the prompt AI created, paste it into a new chat window, and go.

[INTERMEDIATE]

The key to a good prompt is to treat AI as a collaborator, not a search box that gets everything right on the first try. According to MIT’s teaching guidance and OpenAI, better results usually come from clear context, a very specific task, desired format, constraints, examples and refining details in your follow-up. For example, instead of writing, “Help me write a new customer loyalty program,” try this:

Context:

I manage a 12-person specialty retail store. We’ve lost some repeat customers to a competitor’s rewards app over the past years. I want to create my own program, and I have loose notes from a team brainstorm, but nothing is structured yet.

Specific task:

Draft a one-page proposal recommending a point-based loyalty program, including why now, the rough mechanics and all expected costs.

Desired format:

Structure it the following way: problem, recommendation, how it works, pros and cons, and next steps. Keep each section to 3-4 sentences.

Constraints:

Don’t invent specific numbers; flag anywhere I need to fill in real data. Assume the reader is skeptical of new expenses, so frame the problem first, then the solution.

Examples:

I attached a file of a different pitch. Match that level of formality and tone.

After you receive your response, don’t be surprised if you need to refine your initial prompt. You might have left out some details that make the response inaccurate. Take some time to consider all the additional information you’d like to provide, then fill in the details and ask for a revised proposal. Or, if you’d like the information presented in a different format, like an infographic that illustrates the key points, you can ask for that, too.

[ADVANCED INTERMEDIATE]

Customize your response style

If you don’t like the style or manner in which an AI model responds to your prompts, you can personalize it to respond in a style that best suits your needs. Here are some examples of general requests you can make. We recommend you be very specific in your instructions, explaining what you do and don’t want for each preference.

Admit when it does not know or cannot do something.

Avoid logic gaps and unsupported conclusions.

Be direct and intelligent.

Challenge your thinking and assumptions.

Cite sources.

Communicate in a conversational tone.

Link directly to the most useful original or authoritative source.

Look for counterarguments, alternative explanations and important limitations.

Minimize praise.

Prioritize accuracy and truthfulness.

Provide honest feedback.

Test its own biases.

Pro Tip

Not sure if AI can help? Just ask. If you’re working on a task and think, “I wonder if AI can help me with this,” the best way to find out is to simply ask AI.

Try: I’m a [your role] working on [task/project]. What are some specific ways you could help me with this, and what would I need to tell you to get started?

The more context you give (your role, your goals, any constraints), the more useful the suggestion. A vague ask like, “How can you help me?” will get you a generic list. A specific one gives you a real starting point.

How to learn more

Ready to dive in and get the most out of AI? We recommend starting with a free online course. There are classes for professional and personal use, many focusing on specific categories or topics. Here are a few places you can start, but we encourage you to explore what is available because, as with everything related to AI, this list will change.

Anthropic Academy

From the makers of Claude, these courses are great for learning how to use AI safely in the workplace.

DeepLearning.Ai

A combination of free and paid courses that cover a wide range of topics, from basics like prompt engineering to more advanced subjects like building multimodal data pipelines.

Elements of AI

Developed by the University of Helinski, it provides overall training for nontechnical learners.

Grow with Google

Introductory courses for everyday use, educators, students and small-business owners.

HuggingFace

More technical (understanding how to code in Python is recommended), but great if you’d like a deeper dive.

Microsoft Learn

Covers the basics as well as platform-specific training for Azure.

Seeking connection with AI

There are some users who use AI as a companion, seeking connection and emotional support. AI can simulate a relationship, but AI produces answers based on what words or sentences it predicts should come next, not on knowledge of human emotions. It cannot and should not replace a real human relationship. Although AI chatbots can offer a space where users can talk freely without fear of being judged or socially penalized, there are risks, including dependency, distorted expectations of human relationships and bad advice. 

AI chatbots are made to be helpful, agreeable and engaging. They are designed to answer you by predicting what words should come next, often based on what they think the user wants to hear. For people struggling with mental health, using AI chatbots for emotional support can lead to fatal consequences. In July, Meta rolled out a set of guardrails for teens. When parents set up the supervision feature on Instagram, they will receive a notification from Meta if their teen’s conversations indicate they might be in crisis. This initiative might be in response to concerns after news of two children committed suicide and AI interactions were alleged to be a significant contributing factor. However, the risk is not limited to children. ChatGPT is currently being sued for allegedly encouraging a 29-year-old California woman to commit suicide.

AI is not a substitute for professional mental health care. If you are struggling with mental health, it is advised to seek professional help from a human mental health professional.

For more on AI companions, click here.

Addressing your concerns

The United States government has not been proactive in regulating AI as a whole, let alone bias and accuracy (click here for more information). It’s important to use AI responsibly, remembering that there are inaccuracies. You should use it as a tool and not believe the information it provides is fully accurate. 

Reliability, hallucinations and the importance of fact-checking

Hallucination rates are still between 3.1 percent and 19.1 percent, depending on the model, the task and its reasoning configuration. While it’s better than what it was in 2024 (15-45 percent) when hallucinations became a general point of concern, you should still question any claims AI presents to you.

Fact-checking AI’s output should be done carefully every time you use it. Here’s how:

1. Start with your prompt. After you’ve asked your question, add the following sentence: “Please provide links to all the sources you use.” 
2. Depending on your question, you may also want to specify the types of resources the AI can use. 
3. Identify all the facts presented in the response. Then check the sources provided. Are they trustworthy and unbiased (from a highly credible academic institution, a trusted news source, scientific research, etc.) or a stranger’s blog? If you’re presented with news outlets known for their bias, be sure to check them against news outlets on the other side of the political spectrum as well as more neutral ones. Finally, if only a single source was provided, find additional trusted sources on the topic and compare information. 

Another way to double-check your facts is to ask multiple AI models the same question. Then compare their answers and sources.

If you have specific information you’d like to review and want to avoid hallucinations, we recommend using Gemini Notebook. This AI model only works with the documents you upload or websites you provide. If you ask a question that requires information outside of what you provided, Gemini Notebook will not provide an answer without first receiving your permission to check external sources. Here are some examples of how you can use it:

1. Create visual formats like mind maps, slide decks and infographics
2. Cross-reference documents to compare their arguments or find conflicting information
3. Summarize long texts into easy-to-understand audio- or video-overviews  
4. Study specific information
5. Synthesize complex documents and research

For research, we highly recommend Perplexity, an LLM that is more of an answer engine than a chatbot. It’s known for searching in real time (meaning it pulls the most recent data first) and providing inline source citations. And new this year, Perplexity’s “Spaces” feature allows users to organize their research and specific instructions for a project into a single folder, aiding future prompts. As always, even with Perplexity’s stellar reputation, it’s important to fact-check the information provided by heading to each source to ensure it matches what the source said.

AI bias and fairness

AI is not free of human bias because it’s trained on human-created data. This often unintentional bias has been known to amplify prejudices toward marginalized groups.

AI can be sexist. 

A 2024 UNESCO report stated that 44 percent of 133 AI models show gender bias, and in a 2026 cross‑national study of five major AI models, 56 percent of responses about young women labeled them as “fragile,” while AI steered 75 percent of women’s suggested careers toward health and social sciences and gave appearance advice 48 percent more often to women than men. 

AI can be racist.

In 2026, a large‑scale study of four million job applications found that AI hiring tools produced federally actionable adverse impact for 26 percent of Black applicants and 15 percent of Asian applicants, with the same candidates being systematically rejected across employers. 

AI can be ableist.

In 2026, researchers found that major LLM chatbots consistently generated stories portraying people with intellectual disabilities as more dependent, childlike and in need of supervision than people without disabilities. 

AI can be classist.

AI can also discriminate against lower-income groups. A 2026 IFC analysis of AI credit scoring found that thin‑file and low‑income borrowers remain more likely to be denied credit and to receive smaller amounts or worse terms. 

Currently, major AI companies like Google and Eightfold* claim to be working on remedies for these problems. Google says it uses a multi-layered governance approach, which means it is being tested and monitored for bias through multiple groups and checkpoints during development and deployment. Eightfold claims its AI hiring platform uses candidate masking to limit personally identifiable information and is designed to assist recruiters rather than make final rejection decisions. However, these safeguards are company-reported measures and there is no outside confirmation that the systems are free from bias.

Major AI developers—including OpenAI, Anthropic, Google, Microsoft, Meta, Amazon, IBM, and Cohere—say they are working to reduce harmful bias. OpenAI released a general fairness evaluation in 2024 that demonstrated how its model at the time presented gender bias (“Hey Jack! How’s it going?” vs. “Hi Jill! How is your day going?”), and in 2025 it released a more specific political fairness evaluation. Claude’s public-facing constitution includes a set of principles that guide model training, steering Claude away from sexist, racist and toxic outputs. Many other developers, including Google, Meta and Microsoft, publish model cards to demonstrate their capabilities, evaluations, limitations and safety concerns in an effort to be more transparent. But remember that bias is typically context-dependent, which makes it impossible for any AI model to be completely free of bias. 

Fine-tuned open models

While all models can exhibit some form of bias, the extent and type of bias can vary across models and deployment contexts. Fine-tuned open models are less likely to produce biased results because they are built for a narrow, well-defined purpose. Using smaller open models (such as Phi-4, Gemma, Llama, Qwen or Mistral), an organization can fine-tune the AI for a specific function or need by uploading their own vetted data and then testing it across demographic groups. Having this control over the data, along with the ability to update it, is what reduces bias.

Ultimately the way to “fix” bias is by checking yourself. Here are six methods you can use:

1. Ask the AI the same question again, but change the name, race, age and other demographic details in your prompt. If the response is notably different, then there’s a bias.
2. Ask the AI to share what gender and race it assumed when formulating its answer. You could also ask about socioeconomic background and if it assumed any disabilities.
3. In your prompt, tell the AI not to discriminate based on a single or multiple factors. This is best done in addition to the methods listed above.
4. Run a fairness test using tools like IBM AI Fairness 360 and check the model’s scores, rankings or decisions across groups.
5. Test different AI models for yourself using the methods listed above.
6. Give the AI a prompt that should include multiple viewpoints or groups. Then ask a follow-up question that forces the AI to address the missing side and see if it still omits information.

Privacy and surveillance

Collection of data without consent

AI-enabled services may collect personal data without explicit consent, especially when drawn from apps, websites, cameras and other sources. 

Inference of sensitive traits

AI systems can infer sensitive information, including health status, political leanings, relationships or routines from ordinary data.

Data reuse

Depending on the service and its settings, user inputs may be stored, reviewed or used to improve models, so data shared for one purpose may be reused for another.

Security exposure

AI systems can increase the risk of data leaks, hacking and internal misuse because they often process large volumes of sensitive information.

There are ways to minimize your risk of these privacy violations and surveillance when using AI. 

• Never enter sensitive information that you would not like to have shared publicly (passwords, ID, medical or financial data or private documents). 
• Check privacy settings to ensure the data you enter is not used for training purposes; delete old chat history.
• Limit permissions by revoking unnecessary access to photos, contacts, microphone, location and files on AI apps.
• Create a unique password and use multi-factor authentication to prevent your accounts from being compromised.

Systemic concerns: Transparency, accountability, misuse, job impact and environmental impact

Transparency and accountability

Transparency means users know how an AI was built, what data it uses, how it is tested and where it fails. There are concerns that many systems are released with limited disclosure, mainly due to innovation speed. This secrecy makes independent evaluation and public oversight difficult. Due to this lack of transparency, accountability is not taken when things go wrong, as it’s unclear where the responsibility lies, whether it’s the developer, deployer, vendor or user.

Misuse and harm

AI can be used to generate misinformation, impersonation, fraud, surveillance and other harmful content. The U.S. Government Accountability Office flagged the risk that generative AI can spread false information and create or elevate safety concerns. Westchester County passed a law last year to address deceptive business practices and AI scams.

Job impact

There are many concerns about entry-level jobs being replaced by AI, and, according to an April article in The Wall Street Journal, Verizon CEO Dan Schulman says the overall unemployment rate could rise to 20 or 30 percent within the next two to five years. The issue is not mass layoffs; it’s less creation of new job opportunities, according to several Yale professors. This is  especially true for entry-level positions, making it harder for junior workers to gain experience and skills. Click here for more on this topic.

Environmental impact

While AI is being used to help environmental emergencies, such as detecting methane leaks from oil and gas sites and monitoring wildfire spread, there are negative impacts on the environment, too. The main issue is not just the amount of water data centers use; it’s the full environmental footprint, including electricity demand, land use, critical-mineral extraction and e-waste. Data centers impact local and neighboring communities by increasing pressure on local water and power supplies, and they also create heat, noise and light pollution. To explore and hopefully prevent these potential long-term impacts, in July, Governor Hochul enacted a temporary one-year ban on building data centers in New York State. This will give researchers and communities time to understand the impacts and make better, more informed decisions. Click here for more on this topic.

Governor Hochul signed Executive Order No. 62, which established a halt in the development and building of large-scale data centers. This was done to further evaluate the use of energy and water resources that will inevitably affect New Yorkers. Before this order, Governor Hochul and the Department of Public Service (DPS) began Energize NY, which will require data centers to pay more for their energy or supply their own. 

This article was edited by Julie Schwietert Collazo.

This article was published in the September/October 2026 edition of Connect to Northern Westchester.