Unit 1 / 10

AI in Environmental Engineering: Introduction, Boundaries, Responsibility and Ethics

Gains:

  • Risk-based classification of where AI is accelerating and where it is risky in environmental engineering
  • Ability to recognize LLMs' risk of hallucination regarding environmental limit values, standards and resources
  • Ability to validate AI output in laboratory, field and regulatory and apply ethical/privacy principles

Imagine working as an environmental engineer in a common wastewater treatment plant of an organized industrial zone. There is a sudden increase in COD in the outlet water, the business manager says on the phone, "I want the cause and solution as soon as possible." On your desk are laboratory results from the last week, SCADA records, and a chat window you just opened. You ask AI "why did the output COD increase, write down the discharge limit as well?" Within seconds, the model writes in proper Turkish, item by item, the possible reasons and also the sentence "According to the Water Pollution Control Regulation, the COD discharge limit is 120 mg/L." The problem is that this breakpoint may have been model fitted; Depending on your industry, receiving environment and facility type, the actual limit may be completely different. This unit teaches you to manage that very moment—the moment when your engineering responsibility intersects with the rapid but unverified output of AI.

What AI does and doesn't do in environmental engineering

Large language models (LLM) are systems trained on the principle of "predicting the next word" on huge text data. This makes them very strong at language work, but fundamentally weak at fact-checking. The model doesn't "remember" a breakpoint, it produces a number that looks statistically reasonable.

Jobs AI is good at in environmental engineering:

  • Summarizing and drafting long legislation, EIA reports or literature texts
  • Establishing the structure, formula and logic of a calculation (mass balance, emission load, efficiency)
  • Creating a list of possible causes and scenarios for troubleshooting
  • Preparing draft reports, tables, presentations and correspondence
  • Explaining a complex topic (e.g. nitrification chemistry) step by step

Tasks that AI cannot do/is unreliable:

  • Providing an up-to-date and context-specific limit value, standard number or regulation article
  • Access to your facility's actual measurement data (unless you provide it)
  • Replacement for laboratory analysis or field measurement
  • Assume legal/administrative responsibility — signature is yours
Tip: Think of AI like a “senior intern.” Quick, well-read, good draft writer; But you don't send every printout anywhere without checking it yourself. Every number and every reference to legislation he produces is just a hypothesis until you verify it from the source.

Hallucination: fabricated limit value and regulation article

A hallucination is when the model produces unreal information with full confidence. The three most dangerous types of hallucinations in environmental engineering are:

  1. Made-up limit value: Numbers that seem correct but may be wrong, such as "The 24-hour limit for PM10 is 50 µg/m³".
  2. Made-up legislative reference: "According to Article 12/3 of the Regulation..." — the article number and the name of the regulation appear real, but the content is made up.
  3. Fictitious source: A non-existent article, standard number (e.g. "TS EN 12457") or institutional guide.

The table below shows safe and risky answer patterns that the model can give to the same question:

Model output

Risk

what to do

“The discharge limit is exactly 120 mg/L.”

High — exact number, unsourced

Verify from official legislation

"Discharge limits vary depending on the sector and the receiving environment; the relevant regulation table should be consulted."

Low — router

Still verify, but trustworthy

"I cannot verify this value, see official source."

low

ideal behavior

Caution: The more specific and confident a number is written, the more dangerous it is if it is unsourced. A clear value like "120 mg/L" will put you at ease but give false confidence. Do not base your decision on any limit value that is not cited.

Risk-based division of labor

The weight of the outcome of the decision is what determines where to use AI freely and where to use it only for drafting purposes. The following classification is a practical compass:

Quest

Risk level

The role of AI

Mandatory verification

Literature/legislation summary, concept explanation

low

producer, accelerator

Cross-check key claims

Report/correspondence draft, table layout

low

draft writer

Editorial review

Data analysis setup, calculation formula

medium

builder

Input data + manual control of calculation

Process troubleshooting scenario

medium

idea generator

Laboratory + field data

Discharge/emission limit determination

high

Draft only

Official legislation + accredited laboratory

EIA result, "impact negligible" judgment

high

Draft only

Expert + field measurement

Human/environmental health decisions

critical

Support, never decide

Full legislation + expert approval

The rule is simple: the more irreversible the outcome and the more lives/money/laws it affects, the more stringently the AI output is verified. You gain speed at low risk; at high risk AI produces only the first draft, data and legislation make the decision.

Weak prompt / Strong prompt

Two different requests for the same need give rise to completely different risk profiles:

WEAK PROMPT: "What is the wastewater discharge COD limit? Write briefly."Problem: Forces the model to fit a single hard number. Sector, receiving environment, facility type not specified. The output comes without a source. STRONG PROMPT: "Explain which factors (sector, receiving environment, flow rate) the COD discharge limit for textile industry wastewater in Türkiye depends on. DO NOT make up an exact number; instead, explain step by step from which official regulation table and according to which parameters it should be read. If you are not sure, state 'must be verified'." Difference: Strong prompt does not make the model fit a number; It uses it as a GUIDE and leaves the final word to the legislation.

Ethics and data privacy

As an environmental engineer, the data you have is often sensitive: a company's continuous monitoring (CPMS) records, violation history, production capacity are trade secrets. Any data you paste into a public AI tool may be out of your control.

Basic principles:

  • Anonymize: Remove company name, facility coordinates, personal data, contract details from the prompt. Instead of "company X" write "a textile plant".
  • Protect trade secret: Check your corporate data policy before sharing information such as breach data, capacity, cost.
  • Transparency: If you drafted a report with AI, indicate this appropriately in your workflow; Remember that the responsibility is yours.

# A simple anonymization check before putting data in the prompt sensitive_fields = ["company_name", "coordinate", "tc_no", "capacity", "cost"]def prompt_safe_is(text, sensitive): found = [a for a in sensitive if a in text.lower()] return (len(found) == 0, found) example = "company_name: ABC Textile, COD: 850 mg/L, capacity: 5000 m3/day"safe, risky = prompt_safe_mi(example, sensitive_areas)print("Can it be sent:", safe) # Falseprint("To be cleaned first:", risky) # ['company_name', 'capacity']

mini case

A young engineer at an environmental consulting firm asks the AI ​​“what is the waste code and disposal method of waste oil” for an urgent quote. The model confidently returns a 6-digit waste code and the answer "disposal by incineration". The engineer puts this directly into the quote. After delivery, the senior engineer checks: the code provided does not correspond to the actual source and content of the waste; Correct classification can only be made with the analysis report and official waste list. Wrong code means wrong disposal route and risk of administrative sanctions. Lesson: The code the AI ​​gave was a hypothesis; Since it was not verified by analysis and legislation, it could not be the basis for the decision.

Common mistakes

  • Putting the limit value given by AI into the report without asking its source.
  • Considering the reference to "Regulation Article X" reliable without comparing it with the actual text.
  • Pasting confidential information such as company name, tracking data etc. into the public tool.
  • Basing high-risk decision (discharge limit, EIA result) on AI output.
  • Mistaking the model's confident tone for evidence of accuracy.
  • Trying to shift the responsibility to the machine by saying "AI said"; The signature and responsibility lies with the engineer.

In summary

  • AI is an accelerator in environmental engineering: powerful in brief, draft, account editing and scenario generation; It is unreliable when referring to exact limit values ​​and legislation.
  • LLMs produce hallucinations about cutoffs, standards, and sources; precise and unsourced numbers are the highest risk.
  • The division of labor is risk-based: the more irreversible the outcome, the tighter the verification.
  • Every AI output is just a proposal/sketch until validated by laboratory, field measurement and government legislation.
  • Do not provide confidential and personal data to any public means without anonymizing it.
  • AI is a “senior trainee”: it runs, but the engineer carries the signature and responsibility.

Application task

Choose five tasks from your field of study (wastewater, air, waste or EIA) and place each in the risk-based table above: write the task, determine the risk level (low/medium/high/critical), define the role of the AI ​​and the mandatory verification step. Then choose one of these tasks and first deliberately write a "weak prompt", ticking whether there is at least one unsourced reference to limit values ​​or legislation in the answer you receive from the model; Then, for the same need, write a "strong prompt" that prohibits making up numbers and directs you to the source, and compare the two outputs side by side. Finally, describe in one paragraph which government legislation and which laboratory/field data you will use to verify this task.