AI in practice and ethics

Why this matters

A model that scores well in the lab is still far from a usable system. It must run in time on hardware that can fly, work well in every condition it will meet, and have someone accountable when it goes wrong, especially in work affecting people’s lives and privacy, such as searching for victims or monitoring public spaces.

AI on the drone, the ground station or the cloud

Edge AI runs models on the end device, such as the computer on a drone. Popular hardware includes accelerators such as the Raspberry Pi AI HAT+ at 13 and 26 TOPS for the Raspberry Pi 5, and the NVIDIA Jetson Orin Nano Super, rated at 67 TOPS. These TOPS figures are measured under different conditions (the Jetson figure is sparse INT8, for example), so they cannot be compared directly; always time the actual model on the actual board. Models are often converted to an intermediate format such as ONNX and run with ONNX Runtime, and you must check that results after conversion still match the original.

Tasks that must respond within a fraction of a second belong on the drone, tasks that need human confirmation belong at the ground station, and retraining, which needs heavy resources, belongs in the cloud or on a server.

speed = 10.0
fps = 30
budget_ms = 1000 / fps
for label, latency_ms in [("median", 35), ("p95", 80)]:
    print(f"{label:<6} {latency_ms} ms -> drone moves {speed * latency_ms / 1000:.2f} m, "
          f"within frame budget {budget_ms:.1f} ms: {latency_ms <= budget_ms}")
print(f"frames the detector can keep up with at p95: {1000 / 80:.1f} fps")

median 35 ms -> drone moves 0.35 m, within frame budget 33.3 ms: False
p95    80 ms -> drone moves 0.80 m, within frame budget 33.3 ms: False
frames the detector can keep up with at p95: 12.5 fps

Evaluating by group

An overall average can hide groups where the model does badly. Training data with many more daytime than night-time images, or more of some kinds of people than others, makes the model biased against the groups with less data.

import pandas as pd

results = pd.DataFrame({"condition": ["day", "night"], "TP": [90, 24], "FN": [10, 16]})
results["recall"] = results["TP"] / (results["TP"] + results["FN"])
print(results.to_string(index=False))
overall = results["TP"].sum() / (results["TP"].sum() + results["FN"].sum())
print(f"overall recall {overall:.3f}")

condition  TP  FN  recall
      day  90  10     0.9
    night  24  16     0.6
overall recall 0.814

An overall recall of 0.81 looks good, but at night only 60% of victims are found. If real searches often happen at night, the overall figure makes decision makers trust the system too much. These are hypothetical figures for practice. Always report results by the conditions that matter, and state in the data card which conditions the data covers.

People in the decision loop

Levels of automation run from human-in-the-loop, where the system proposes and a person decides, through human-on-the-loop, where the system acts but a person watches and can stop it, to full autonomy. Decisions with severe or irreversible consequences should rest with people. AI output from screening disaster imagery should therefore be a list for checking, not a verdict on real conditions, as the knowledge unit on AI for screening disaster imagery explains.

AI ethics and governance frameworks

• The NIST AI Risk Management Framework 1.0 (January 2023) manages AI risk through four functions: Govern (policy, roles, accountability), Map (understand context and risks), Measure (measure and test) and Manage (prioritise and treat risks). NIST states that the framework is being revised.
• UNESCO’s Recommendation on the Ethics of AI (November 2021) sets four core values, such as human rights and human dignity, and ten principles, such as proportionality and do no harm, privacy, transparency and explainability, human oversight, and fairness.
• The EU AI Act (Regulation (EU) 2024/1689) entered into force on 1 August 2024 and applies in stages. Bans on certain AI systems began in February 2025, and the European Commission’s page states that the rules for high-risk systems have been postponed to 2027–2028. The timeline can still change, so check the latest version before citing it. The law applies to providers placing systems on the EU market.
• Thailand: ETDA and the Ministry of Digital Economy and Society held a public consultation on draft principles for an AI law in June 2025. At the time of writing, no enacted law was found, so check the latest status. Personal data in images and training data falls under the Personal Data Protection Act B.E. 2562 (2019), covered in UAT 313 module 5.

Class activity

Common mistakes

Summary

• Choose where to process by the response time needed, and time the real model on the real board; distance moved during processing is $v\times t$
• Evaluate results by group to find bias hidden in averages
• Severe-consequence tasks need people in the decision loop; AI output is a list for checking
• Use the NIST AI RMF and UNESCO’s principles as frameworks, and check the current status of the EU AI Act, Thai AI law and the PDPA

Check your understanding

1. A drone flies at 15 m/s and the detector takes 60 ms. How far does the drone move before the result arrives?
2. What is the time budget per frame for a 25 fps camera?
3. Day: TP 45, FN 5. Night: TP 12, FN 8. What is the recall of each group?
4. What are the four functions of the NIST AI RMF?
5. Why can TOPS figures from two boards not be compared directly?