Most of us first learned ChatGPT as a conversation. Chat feels like it is asking: What can I answer for you today? That is useful. ChatGPT Work feels different. Work feels like it is asking: What can I take off your plate? You are not just asking for an answer. You are giving ChatGPT an assignment. You define the outcome. You give it context. You let it use connected apps where available. Then you review the result before anything happens. That last part matters. OpenAI currently describes Work as the place for longer, multi-step tasks and finished deliverables, and apps can connect ChatGPT to external tools, information, and actions depending on your plan, workspace, permissions, and settings. (OpenAI Help Center⁠) Start With a Safety Boundary When you are first learning ChatGPT Work, I recommend adding this line to your prompt: Review and prepare the work, but do not apply, submit, send, message, delete, archive, move, update, schedule, cancel, publish, or modify anything without my explicit approval. That may feel like overkill. It is not. Connected apps make ChatGPT Work more useful because the information may already live in your email, calendar, files, or other tools. But useful does not mean uncontrolled. My simple rule is: Analyze first. Recommend next. Act only after approval. Practice This: Job Search Review For the first practice, let’s use a job search review. This is a good starting point because it is useful, but still easy to inspect. We are not asking ChatGPT to apply for jobs. We are not asking it to message anyone. We are not asking it to upload anything. We are asking it to review what it can access, organize the opportunities, and prepare a recommendation for us. Depending on what you have connected, ChatGPT Work may be able to review things like: - Job alert emails - Recruiter messages - Saved job descriptions - Career notes - A resume or professional summary - Relevant files in connected storage

When AI Looks Almost Human, “Almost” Becomes the Problem. Have you ever watched an AI-generated person who looked realistic, but something still felt wrong? The face looked human. The voice sounded human. The movements were close to human. But your brain still whispered: That is not a real person. That uncomfortable reaction is known as the uncanny valley. 🌀 What Is the Uncanny Valley? As robots, avatars, and digital humans become more realistic, people generally respond to them more positively. But there is a point where something looks almost human without being convincing enough. At that point, our comfort level can suddenly drop. Small imperfections become strangely noticeable: - Eyes that do not move naturally - Facial expressions that arrive a fraction too late - Lips that do not perfectly match the words - Skin that looks too smooth - A voice that lacks natural emotion - Body movements that feel slightly mechanical A cartoon character does not bother us because we know it is not human. We are not expecting perfect realism. An AI avatar that looks 98 percent human creates a different expectation. Our brains become extremely sensitive to the missing 2 percent. ⚠️ Why This Matters for AI: AI-generated video, voices, virtual assistants, and digital employees are improving quickly. The goal, however, should not always be to make AI indistinguishable from a human. Sometimes a clearly artificial character feels more trustworthy than a digital person pretending to be real. This creates an important design question: - Should an AI experience try to look completely human, or should it be honest about being AI? There is no universal answer. It depends on the audience, purpose, and level of trust required. The uncanny valley reminds us that technical realism and human comfort are not the same thing. As AI becomes more humanlike, the smallest details may determine whether people feel connected, curious, uncomfortable, or deceived. Share in the comments:

In Part 4, we talked about dialing model capability up or down. 📌 Now let’s practice it. My shorthand is: Find the floor, then move back up one level. That means you reduce the reasoning or effort level until you can see the model starting to miss the mark. Then you move back up to the lowest setting that still produces the quality you need. That gives you a practical balance between: - Output quality - Speed - Cost - Available usage or capacity - The amount of correction you have to do yourself Practice This: - Choose an assignment you already understand well. - Do not use something completely unfamiliar. - You need to be able to recognize when the output gets weaker. Good practice assignments include: - A business proposal - A project plan - An app framework - A book chapter outline - A training lesson - A process improvement plan ✅Step 1: - Use the model map from Part 3 to choose your product and model. - Start with the model you believe is appropriate for the work. ✅Step 2: - Run the assignment using a relatively high reasoning or effort setting. - Your goal is to create the first strong version of the plan, framework, proposal, or structure. - Save the result. This becomes your quality baseline. ✅Step 3: Once the main thinking is complete, reduce the reasoning or effort level one step. For example: High reasoning → Medium reasoning Or: High effort → Medium effort If the work is now very clear, and the product allows it, you can test reducing it two levels. ✅Step 4: Ask the model to continue the work. For example: - Expand one section - Rewrite part of the proposal - Create a summary - Draft an email from the plan - Turn the framework into a checklist - Create social posts from the article - Refine the tone - Format the output for easier reading This is where the test begins. You are no longer asking the model to do the hardest thinking from scratch. You are asking it to continue from an established structure.