← Prev Next →

AI Presentation 2026 - LLM AI Deficits

How the models fail, who absorbs the damage, and who answers for it

1. Where the Damage Lands

This page draws on Global_Liabilities_of_LLM_AIs.docx, which sorts the liabilities into eight areas and matches them against the benefit list on the previous page. One pattern runs through all eight. The same low cost that puts a tutor or a translator in everyone’s hands also puts a forger, a scammer, and a confident wrong answer there. Scale cuts both ways, and the people who absorb the damage are rarely the people who chose to deploy the model.

2. Misinformation and Trust

The model writes fluent text whether or not the content is true, and it writes as much of it as anyone asks for.
details How LLM output erodes trust An LLM generates fluent text at near-zero cost. That output takes three forms: hallucinations, deliberate fabrication, and sheer volume. All three reach the information environment, where two opposite failures follow: false content is believed, and true content is dismissed as AI-generated. Trust erodes in both directions. LLM Generates fluent text at near-zero cost Output Hallucinations Confident statements that are false, worded exactly like correct ones Fabrication Propaganda, scams, fake reviews, and forged correspondence Volume More content than any reader or moderator can check Information environment What readers encounter and believe False content is believed Readers act on text that no one checked True content is dismissed Anyone waves off a real document as AI-generated Trust erodes in both directions

3. Economic Disruption

The work the model does well is work people are paid for today.
details How LLM capability reaches the labor market A model performs paid knowledge work. Three effects follow: displacement of roles, downward pressure on wages, and an entry cost only large firms can meet. All three act on the labor market and on who owns the tool, producing two results: workers in routine knowledge jobs absorb the loss first, and ownership concentrates in a few companies. Model does paid work Writing, support, translation, programming, administration Displacement Roles in writing, support, translation, and admin shrink or disappear Wage pressure The same work pays less once a model writes the first draft Entry cost Only firms that can buy the computing hardware can compete Labor market and ownership Who does the work, and who owns the tool Routine work absorbs it Routine knowledge jobs and outsourcing economies first Ownership concentrates A few companies hold the hardware and the profit The gain and the loss land on different people

4. Inequality and Access

The tool reaches the connected first, and it serves them best.
details Who the models reach, and how well they serve them Using a model requires means. Three filters apply: a connectivity gap, weaker performance in lower-resource languages, and training data drawn mostly from Western and English-language sources. Together they decide who the model serves well, which leaves two results: the already-connected gain first, and one set of perspectives answers everyone. Using a model takes means A connection, a device, and often a subscription Connectivity gap Fast internet, a current device, and a paid tier decide who gets in Language coverage Models answer worse in lower-resource languages, so their speakers get less Data skew Training text comes mostly from Western, English- language sources Who the model serves well Access, language, and whose text it learned from The connected gain first The gap widens between those with means and without Perspectives narrow One worldview answers questions from everywhere The tool reaches the connected first, and serves them best

5. Bias and Fairness

A model trained on human text repeats what that text contains, including what it gets wrong about people.
details How bias in training text reaches people A model trained on human text absorbs the patterns in that text. Three consequences follow: it repeats stereotypes, institutions apply it to decisions about hiring, lending, policing, and benefits, and billions of people receive advice from the same few models. The results are unfair outcomes at scale and a flattening of culture. Trained on human text The model absorbs the patterns that text already contains Stereotypes repeat The model restates what its sources assumed about people Applied to decisions Hiring, lending, policing, and benefits reviews use the output One source, many users Billions of people read writing and advice from the same few models Decisions and culture What institutions decide, and how people write Unfair outcomes at scale One biased pattern decides thousands of cases Culture flattens Advice and prose converge on one house style Bias in the data becomes bias in the decision

6. Security and Misuse

The model lowers the effort an attack takes, which raises the number of people who can mount one.
details How models lower the cost of an attack A model lowers the effort an attack takes. Three effects follow: fraud becomes personalized and cheap, writing malware demands less skill, and people seeking dangerous technical information get further than they would alone. The attack surface grows in two directions: more people can attack, and each attacker reaches more targets. Effort drops The model does the writing and the technical work Personalized fraud Phishing and social engineering, each message written for its target Lower skill barrier Writing malware and mounting cyberattacks takes less expertise Dangerous know-how Seekers get further than they would alone, which is why firms add safeguards Attack surface Who can attack, and how many they can reach More attackers Skills that gated the work no longer gate it More targets each One operator sends thousands of tailored lures Effort falls, so volume rises

7. Privacy and Data Rights

Text entered into a prompt leaves the user’s control, and text used in training mostly arrived without permission.
details How text leaves its owner's control Text reaches a model from two directions: what users type into prompts and what companies gather for training. Prompt text gets stored and sometimes exposed, training text arrives without clear consent, and a model that reads quickly makes profiling easier. Two results follow: exposure of private material and unsettled ownership. Text enters the model From what users type and from what firms collect Prompt data Personal and company text gets stored, and sometimes exposed Training data Copyrighted and personal material, gathered without clear consent Reading at scale A model works through volumes of text that no staff could review Control over your own text Who holds it, who may use it, and for what Exposure Private material surfaces, and profiling gets cheaper Ownership unsettled Courts have not decided who owns what Text leaves its owner’s hands from both directions

8. Environmental and Resource Costs

The models run in buildings that consume power and water in the places that host them.
details What running the models costs physically The models run in data centers. Those buildings draw electricity for training and for answering questions, consume water for cooling, and fill with hardware that must be manufactured and later discarded. The load falls on host communities and the local grid, so the cost is local while the benefit is global, and the waste outlives the model. Models run in buildings Data centers sited for power, not for proximity to users Electricity Training a model and answering each question both draw power Water and grid Cooling consumes water, and the load strains the local supply Hardware Manufacturing the chips carries its own cost, and they end as e-waste Host communities and the grid The places that supply the power and the water Local cost, global benefit Neighbors carry the load for users everywhere Waste outlives the model Retired hardware remains after the model is replaced The computing has a physical address

9. Human and Social Effects

A tool that answers for you can take the practice away from you.
details What people stop doing when the model does it The model answers in place of the user. Skills that go unexercised weaken, students and professionals submit work they do not fully understand, and some users form emotional dependence on a chatbot. Two results follow: capability erodes, and human contact narrows. The model answers for you Drafting, reasoning, and company on demand Skills go unused Writing, research, and critical thinking need practice to hold Unexamined work Students and professionals submit output they do not fully understand Dependence Some users rely on a chatbot for company, not just for answers What people stop practicing The work, the judgment, and the conversation Capability erodes The skill to check the output fades with it Contact narrows Lonely and vulnerable users see fewer people Convenience now, capability later

10. Governance and Accountability

When a model causes harm, no one has settled who answers for it.
details Why harm from a model is hard to assign When a model causes harm, responsibility sits somewhere between the developer, the company that deployed it, and the user. Regulation varies by country and trails the technology, and auditing is hard because the builders cannot fully explain a given answer. The result is an accountability gap: no one is liable by default, and an audit cannot settle it. A model causes harm Bad advice, a wrong decision, a loss someone has to carry Three candidates The developer, the firm that deployed it, or the user who acted on it Patchwork rules Regulation differs by country and arrives after the technology Opaque systems The builders cannot fully explain why a model gave the answer it gave Accountability gap No settled answer to who owes the injured party Liability falls through Each party points to the other two Audits fall short A review cannot recover why the output appeared Harm arrives before the rules do

11. What Reduces the Risk

The source document ends on a qualified note: safeguards, regulation, transparency, and human oversight each cut these risks. How much they cut depends on how the technology is deployed, which returns the question to the people doing the deploying.